This PDF file contains the front matter associated with SPIE Proceedings Volume 12754, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

The current traditional multi-scale visual planning model mainly relies on existing a priori maps to realize visual planning, which leads to poor planning results due to the lack of establishment of reward functions. In this regard, a digital multiscale visual planning model for the geographic landscape pattern of smart parks is proposed. The AC algorithm is used to describe the observed state and update the planning strategy directly by maximizing the average return, and establish the planning reward function. In the experiments, the planning performance of the proposed model is verified. The analysis of the experimental results shows that the visual planning model constructed by the proposed method possesses a lower collision rate and the model has a better planning performance.

In order to solve the problem of insufficient learning ability of mutual representation of two modal information in existing visual question answer models, the Global-Local Attention Network (GLAN) approach to visual question answer model with global-local features is proposed by analyzing the reasons why the model cannot distinguish the relationship between similar targets in images and the reasons for the model's reliance on text information. To supplement the missing local information in the global features, the model is designed with a convolution guided attention module to enhance the image representation capability and increase the bias of image information in the model. Experiments are conducted on VQAv2.0 and GQA datasets, and the accuracy of the proposed model is improved by 0.60% and 0.38%, respectively, compared with Baseline, which can effectively improve the inference ability of the model.

This paper proposes a novel method for detecting object sizes in images. Traditional methods have limitations due to image quality, which affects detection accuracy. To address this problem, a new approach is proposed that combines Zernike moments and deep learning models. This approach combines the strengths of both methods and overcomes the limitations of traditional approaches. The proposed approach improves the accuracy of object size detection and is robust to image noise and distortion. Experimental results demonstrate the effectiveness of this approach, which outperforms existing methods in terms of accuracy and robustness.

Most of the land surface water in the plateau region is frozen. Due to the huge difference in the dielectric constant of water and ice, this paper generated a ten-year surface permafrost change map from 1992 to 2001 based on SSM/I passive microwave remote sensing data, and the results showed that from 1992 to 2001, the annual average frozen area in Tibet showed a downward trend, which also verified the general trend of global warming. According to the comparison between the images generated by the two algorithms and the actual situation, it can be seen that the freeze-thaw discriminant algorithm has better accuracy than the double-index algorithm under the condition of ascending orbit. The results of this study can have important reference value for analyzing the changes of permafrost in Tibet for a long time series and selecting the appropriate soil freeze-thaw discrimination algorithm, and also provide an SSM/I sensor image processing method.

With the rise of the take-out food industry and bicycle sharing, the frequency of non-motorized vehicle using has increased significantly, and the ensuing safety hazards have become increasingly serious. Since motor vehicles are always subject to traffic violations by non-motorized vehicles, it is important to classify non-motorized vehicles for optimal handling in emergency situations. Due to the uniqueness of the target information, the micro-Doppler effect is increasingly used to distinguish between different targets. In this paper, we propose a non-motorized vehicle classification method based on the micro-Doppler effect and image processing. Firstly, the relationship model between ground non-motorized vehicles and radar is established, the expression of the echo signal is derived, and the simulation analysis is carried out according to the set parameters, the differences between the non-motorized time-domain frequency domain diagrams and the joint time-frequency diagrams were investigated. Secondly, the classification features of the car hub echo signal are introduced, and the feasibility of these features as classification criteria is verified by simulations. Then, the feature extraction methods based on image processing are introduced, and the extraction results of different edge detection operators are compared and analyzed. Finally, the classification method of non-motorized vehicles based on these three features is introduced.

Existing visual simultaneous localization and mapping (SLAM) systems usually assume that the environment is static, therefore when many mature visual SLAM algorithms are applied to dynamic scenes, the pose errors estimated by the system will increase sharply due to the interference of dynamic objects. To address this problem, this paper proposes a robust visual SLAM system based on ORB-SLAM2 using static features and realizes a two-stage tracking from coarse to fine to improve the system localization accuracy in a dynamic environment. The proposed method uses the yolov5-seg network to implement a semantic thread, then combines semantic prior information and geometric information to accurately extract dynamic feature points, and realizes pose optimization based on the static weight of feature points. In addition, an improved keyframe selection strategy is implemented. Experiments on the public TUM dynamic datasets show that the proposed algorithm performs well on most dynamic datasets, has a great improvement compared with the ORB-SLAM2, and achieves the same accuracy as other advanced dynamic SLAM systems.

Object detection is the key task for many Unmanned Aerial Vehicles (UAVs) tasks. Compared with images captured from ground-based angles, images captured by UAVs are characterized by large differences in object scales, complex backgrounds, and dense objects. Common object detection models on UAVs platforms are not very accurate or effective. To solve the above problems, in our work, we propose a real-time object detector based on improved YOLOv5N for UAVs remote sensing images. Firstly, we introduce GhostNet to replace the standard convolutional layer in the feature extraction and feature fusion network of the original YOLOv5N, thus reducing the computational overhead of the detector. Secondly, the last layer of the backbone network is substituted by a Separable Visual Transformer (SepViT) block to enhance the connectivity of the backbone with global information. Finally, the Efficient Channel- Coordinate Attention (ECCA) module, realized by combining the Efficient Channel Attention (ECA) network and the Coordinate Attention (CA) module, has also been added to YOLOv5N, by combining channel information and spatial information to highlight information that helps detect small objects. The evaluation results based on the VisDrone2019-DET dataset show that the dataset is widely used for UAV remote sensing image evaluation, and the proposed model achieves 32.9% average accuracy (mAP50), which is 6.8% better than YOLOv5N, 15.6% reduction in model computation, 6.7% reduction in inference time, and both detection accuracy and detection speed, it is suitable for deployment on UAVs for real-time object detection tasks.

Aiming at the problem of special shape of graptolites, insufficient training samples and manual identification of graptolites consumes a lot of time and energy, a new Graptolite image classification based on feature transfer and mixup data enhancement is proposed. First, we expand graptolite dataset by using traditional data enhancement and mixup to generate mixed samples, then use the deep learning model pretrained on the large-scale ImageNet dataset to construct a new transfer learning model through feature transfer. Finally, tenfold cross validation method is used to test the recognition accuracy of the model. The experimental results show that the classification accuracy of the proposed method on the graptolite dataset is 88.65%. Compared with the traditional data enhancement method and the pretrained model trained by ImageNet directly, the classification accuracy of graptolites is improved by 1.19%, which shows that the method has better recognition effect.

This paper establishes a dual stereo vision system and uses Zhang's calibration method to calculate the homography matrix and various parameters of the binocular camera to achieve system correction. The effect of different processing methods on the UAV image in a low-altitude urban background is analyzed. A multistage filter fusion algorithm combined with various filtering and segmentation methods is designed to realize the background suppression of the low-altitude urban background UAV image and drone target extraction. An improved image processing method is used to suppress the background of a single frame image of a video sequence. The image information other than the UAV is regarded as noise or background. The noise and background are subtracted to finally get the UAV image. This research provides a framework for UAV detection in low-altitude airspace, which can support urban airspace management.

Based on PyTorch framework, this paper studies the network structure of classical convolutional neural network VGG16 and the process of input image changing in the network. Using VGG16 to train the small dataset CIFAR10 in various ways, it is found that using the Kaiming normal distribution to initialize the weight for training is far less effective than transfer learning (using the training weight of VGG16 on the ImageNet dataset as the initial value of the weight). The input image of VGG16 can be of any size, but after the image of CIFAR10 is resized to 224×224, the accuracy rate on the test set is improved by 6 percentage points.

This paper mainly applies fuzzy logic to the edge detection of handwritten images in English. Image edge detection is an important part of image processing. Through image edge detection, the amount of data can be greatly reduced, and the information that can be considered irrelevant can be eliminated, and the important structural attributes of the image are retained. It uses six traditional operators to detect the edge of English handwritten images: canny, log, sobel, sprewitt, zerocross, and roberts. By comparing the edge detection results of the two kinds of experiments, we can find that the canny operator has the best edge detection effect among the traditional operators. In the fuzzy logic experiment, with the growth of 𝐒_𝐱 and 𝐒_𝐲, the edge detection effect gradually weakened. In this paper, the experiment with 𝐒_𝐱 = 𝟎. 𝟎𝟓 and 𝐒_𝐲 = 𝟎. 𝟎𝟓 has the best edge detection results. In conclusion, the edge detection effect of English handwritten images based on fuzzy logic is better than that of English handwritten images based on traditional methods.

Image classification technology is at the heart of many complex computer vision applications, including object tracking, video categorization, and action identification. In the area of picture classification, data-driven supervised learning has had considerable success, but these model's capacity to generalize is highly reliant on a huge amount of labeled data. Not all job contexts, though, may quickly acquire a sizable number of labeled dataset samples. This article builds a deep learning approach appropriate for few-sample image classification tasks utilizing Meta-Learning in order to address the issue of image classification tasks in scenarios with few samples. Finally, to confirm the success of the approach, comparative experiments are run on the handwritten letter data set from Omniglot.

To address the problems of low segmentation accuracy and machine complexness of ancient pulse coupled neural network (PCNN) in medical image process, a Converged-FCMSPCNN (CFC-MSPCNN) model is projected. Compared with earlier PCNN models, this model additionally optimizes and enhances the synaptic weight matrix, link strength and dynamic threshold, simplifies the parameter settings and reduces the quantity of iterations. In addition, we add a balance parameter Q to regulate the dynamic threshold to improve the model's control over neuronal image processing. Through relevant experiments, we demonstrate that our algorithm has higher results compared with alternative algorithms to accurately section carcinoma lots and considerably reduces the randomness and unpredictability of firing neurons.

To address the problems of accurate region localization and effective data enhancement, this paper proposes a bilinear convolutional neural network model based on multiple branches and multiple scales. The method uses three branching networks to supervise the image recognition process, which enables the network model to learn fine-grained image features of different scales and different parts. The experimental results show that the bilinear convolutional neural network model based on multi-branch and multi-scale proposed in this paper has practical application value.

As an important feature element in remote sensing images, road network extraction in remote sensing images has become a research hotspot for scholars at home and abroad. With the development of deep learning, road extraction based on the convolutional neural network has achieved certain results. However, due to the complexity of the road environment, it is still difficult to quickly and accurately separate the road content from the surrounding land cover from the high-resolution remote sensing images. In this paper, we make improvements based on the LinkNet network to address the above problems: firstly, in the coding layer part, we design a new fused convolutional module to replace the original residual module in the coding layer, and use multiple small convolutional kernel networks to stack instead of large convolutional kernel networks to improve the operation efficiency of the model; secondly, we introduce an attention mechanism to enhance the focus on road regions and suppress the influence of background regions on road target The second, the attention mechanism is introduced to enhance the attention to the road region and suppress the influence of background region on the road target segmentation.

The feature map is often used for describing the characteristics of key points in image registration applications. Due to feature information being sensitive to rotation and noise, the values of calculation methods for various feature maps would change in different rotation or noise conditions. Consequently, it is unfavorable to obtain the stable descriptor for the same feature point in different images. Specifically, speckle noise damages the synthetic aperture radar (SAR) image seriously, which makes SAR image registration more difficult. In this paper, a novel feature map, which is robust to speckle noise, is proposed. A fully size-adaptive sliding window method is embedded into the ratio of exponentially weighted average method for its noise robustness improvement. Then, a series of Monte Carlo experiments compare the rotation and speckle stabilities among various feature map methods comprehensively. The experimental results reflect that the proposed feature map method is stable while speckle noise increases, and it has less mean orientation difference than others. Moreover, the data of experimental results can be used to support many other analyses on feature maps.

Aircraft detection in synthetic aperture radar (SAR) images plays an essential role in both civil and military fields. However, due to the special imaging mechanism of SAR images, the aircraft annotating process is easily affected by interferences and noises in the background, leading to a high labeling cost. As most object detection networks are trained in a supervised manner, a serious problem of applying them to SAR aircraft detection tasks is the insufficient training data. To address this problem, we propose an unsupervised domain adaption method for the training of SAR aircraft detectors. First, we propose to transfer knowledge from optical aerial images in which aircraft annotations are easier to obtain. By adopting an image-level domain adaption, the target information in optical images can be utilized for the training of SAR aircraft detectors. Then, CycleGAN is adopted to overcome the discrepancy between optical and SAR domains by image-style translation. To evaluate the effectiveness of the proposed method, we build up an optical-to-SAR aircraft detection dataset (O2SADD) based on existing public datasets. Experiments on O2SADD indicate that the proposed method can significantly improve the performance of SAR aircraft detectors without manually annotating aircraft in SAR images.

Echocardiography segmentation is important to quantify cardiac function. Although the segmentation algorithms have achieved considerable performance in image segmentation, the significance of those state-of-the-art (SOTA) segmentation algorithms for clinical diagnosis is not clear. Therefore, this paper comprehensively discussed the significance of the SOTA segmentation algorithm for clinical diagnosis by using segmentation evaluation metrics and diagnostic evaluation metrics. Firstly, four different SOTA segmentation algorithms were used to segment the 2D echocardiography, and the segmentation results were evaluated by using the Dice similarity coefficient (DSC) and Hausdorff distance (HD) which were defined as the segmentation evaluation metrics). Secondly, on the basis of the segmented areas, four diagnose-related image features were extracted, including lower wall thickness (LWT), front wall thickness (FWT), left ventricular transverse diameter (LVTD), and left atrial thickness (LAT). Then, the correlation and difference of those features extracted from the segmented areas and ground truth were statistically analyzed, which were defined as diagnostic evaluation metrics. The results showed that SwinUNet model had the best segmentation results in terms of segmentation evaluation metrics. Among them, SwinUNet's mHD value reached (16.95±1.38)mm and its mDSC value reached (90.16±2.78)%. However, for diagnostic evaluation metrics, there was a significant difference in the LVTD (P<0.05) and LWT (P<0.05) extracted by SwinUNet segmented areas compared with the ground truth. In contrast, there was no significant difference in the LVTD (P=0.026) and FWT (P=0.912) extracted by TransUNet segmented areas. Similarly, LAT (P=0.959) and LWT (P=0.508) from ScaleFormer segmented areas also correlated with those from the ground truth. Four diagnose-related image features (LVTD, LAT, LWT, FWT) extracted from TransUNet segmented areas had the highest correlation with those from the ground truth, and their Pearson correlation coefficient reached 0.95, 0.86, 0.66 and 0.71, respectively. The experimental results showed that the segmentation algorithm with higher segmentation evaluation scores was not necessarily suitable for clinical diagnosis.

With the development of Transformer and its derivatives over the last two years, several research have integrated Transformer with CNN or in instead of CNN to advance medical image segmentation. Although they often create acceptable feature maps for large organs, segmentation accuracy for small organs is less than satisfactory. Transformer excels at global context, but it displays limits in capturing fine-grained features, especially in medical areas, because convolution is unable to describe long-term correlations prevalent in images. This is because local information modeling lacks a spatial inductive bias. Currently available Transformer-based segmentation networks are not often optimized for this issue, we therefore use transformer as the backbone and provide a medical image segmentation network with deformation attention. The model uses attention mechanism to boost the impact of feature maps to address the issue of low accuracy of tiny organ segmentation in multi-organ segmentation tasks. On the Synapse dataset, our model has so far produced SOTA results.

Gastrointestinal diseases are one of the important causes of gastrointestinal tumors, and also one of the serious public health problems that human beings are currently facing. Gastrointestinal inflammation and polyp lesions are the main ways to cause gastrointestinal cancer. Therefore, timely detection of gastrointestinal polyps and inflammation and other lesions through gastrointestinal endoscopy can effectively prevent corresponding cancer. In recent years, deep learning technology has played a huge role in computer-aided gastrointestinal endoscopic diagnosis, especially the convolutional neural network can recognize gastrointestinal lesions. This paper uses the public gastrointestinal endoscopy data set Kvasir (including 7 categories, 1000 images for each category), proposed and designed CBAM-AlexNet deep learning model, in which CBAM is the attention mechanism part, including channel attention and spatial attention, AlexNet is a deep learning network. The original features are input into the channel attention mechanism in CBAM to extract channel features, and then the output channel features are input into the spatial attention mechanism to extract spatial features, and finally the output spatial features are input into the AlexNet network for feature extraction and gastrointestinal disease data. The average accuracy rate, average accuracy rate, average recall rate and average F1Score of all categories of CBAM-AlexNet model were 91.36%, 91.34%, 91.39% and 0.913, respectively, and achieved good classification effect in seven categories of gastrointestinal endoscopy data, showing the superiority of the model. CBAM-AlexNet model has high specificity and sensitivity for the recognition of gastrointestinal polyps, which can assist doctors in the clinical diagnosis of gastrointestinal polyps quickly and reduce the probability of misdiagnosis, Some recognition errors caused by quality issues such as overexposure or underexposure of images will be improved in future work.

Image segmentation of Renal Cell Carcinoma is a crucial prerequisite for pathologists to diagnose the disease and implement treatment. In various previous medical image segmentation tasks, convolutional neural networks(CNNs) have been widely used and have become a practical benchmark with significant success. However, the inherently local nature of the convolution operation leads to its inability to balance long-range relations and local dependencies in the modeling process, resulting in redundancy-deepening networks and the loss of local details. With its innate global self-attention mechanism, transformers were designed for sequence-to-sequence prediction to better extract global information. However, on account of the lack of low-level details, it may cause limited localization abilities. In this paper, we propose a parallel architecture based on Transformers and CNNs, combining Transformers and CNNs in a parallel manner to obtain global and local information by utilizing location-encoded Transformers model and local convolution neural networks, respectively. So that global contextual information can be captured more efficiently while maintaining a strong grasp of low-level spatial detailed information for Renal Cell Carcinoma image segmentation. Through a series of comparative experiments and extension experiments, the effectiveness and progressiveness of our method were validated, demonstrating remarkable potential in the image segmentation task of Renal Cell Carcinoma.

In deep learning-based multi-view stereo methods, the accuracy of feature extraction and the effectiveness of cost volume generation directly affect the reconstruction accuracy. To address the issues of feature extraction and depth map accuracy in current methods, this paper proposes an improved 3D reconstruction network based on a multi-scale cost volume. Firstly, in order to obtain more accurate feature information and improve feature matching, an attention module is incorporated into the feature pyramid extraction network to output three different scale feature maps, and depth estimation is performed using a coarse-to-fine cascade structure. Secondly, to alleviate the problem of over-smoothed edges caused by the regularization process due to the large receptive field and depth map upsampling, the refined upsampled depth map is guided by the feature maps of the next stage with more detailed information, thus enhancing the edge detail information of the initial depth map. Experiments on the DTU dataset show that the proposed method achieves an accuracy of 0.323mm and a completeness of 0.361mm, with a 6.2% improvement in reconstruction completeness compared to the original model, effectively improving the 3D reconstruction performance and verifying the effectiveness of the proposed method.

For the problems of slow speed, low efficiency, and labor intensity of manual measurement of crankshaft structure size of air compressor, a machine vision-based method for measuring crankshaft size of air compressor was proposed. Firstly, a bilateral retention filter was used to filter the grayscale image to ensure the crankshaft edge while smoothing the image. Then the smoothed image was binarized by an automatic threshold segmentation algorithm to segment the crankshaft and the background region. Finally, the sub-pixel edge contours of the image were extracted using canny as the localization function, and the edge contours of the crankshaft were obtained after segmentation, screening, and combination. Based on the extracted contour lines, the edge line regions were drawn, and the pixel distance between the two linear regions was calculated to obtain the diameter size of the crankshaft. The experimental results show that the accuracy of the diameter dimension calculated by this algorithm can reach 99.898% compared with the measured dimension, and the error was less than 0.03mm.

The logistics industry is an important pillar of national economic development. In the logistics and warehousing industry, conveyor belts are one of the important equipment for transporting goods. However, long-term operation can easily lead to various types of defects such as belt breakage, cracking, and wear, which seriously affect logistics operations. Therefore, this paper proposes a defect detection solution for logistics conveyor belts based on machine vision technology. By using machine vision technology to filter, segment, and edge detect images, the corresponding feature parameters are extracted and input into a BP neural network for learning, so as to identify different types of defects. This study uses this solution to process and analyze the surface images of conveyor belts samples through machine vision technology, which can accurately identify different types of defects and their locations, achieving automated defect detection of conveyor belt surfaces and improving logistics operation efficiency and safety.

Medical endoscopes enter the human body cavity through physiological or surgical openings to observe the pathological changes of the tissue structures inside the cavity, helping doctors diagnose diseases. Medical endoscopes are also auxiliary medical devices used in minimally invasive surgical techniques. They are widely used in clinical practice, and their metrological performance directly affects the diagnosis and treatment of diseases by doctors. They are classified as Class II moderately risky medical devices and Class III higher risk medical devices in the "Medical Device Classification Catalog." Strict control and management or special measures must be taken to ensure their safety and effectiveness. Based on this, this article designs a portable and field-calibrated medical endoscope calibration device. Through this calibration device, the metrological performance of the medical endoscope can be objectively evaluated to ensure the accuracy and reliability of the medical endoscope's measurements, and achieve effective traceability.

This paper aims to solve the problem of deploying face liveness detection methods on mobile terminals with low computing resources. A lightweight face liveness detection method based on 32 × 24 thermal imaging images and ShuffleNet is proposed. This method can be deployed on mobile devices with low computing resources on the premise of ensuring detection accuracy. In addition, this paper also proposes an optimization strategy, which can completely eliminate the situation that the fake face is misidentified as a real face, and at the same time control the probability of the real face misidentified as a fake face at a very low level. The experimental results show that the proposed method has high practical value.

The automatic diagnosis of depression using deep learning has recently shown significant progress, and 3DCNN has been applied to assessing depression levels from videos. There are two problems to be encountered applying 3DCNN, including insufficient data to train the model and an insufficient diversity of features extracted on a single scale. An approach is presented in this paper to address these issues. Firstly, we augmented the raw data by adding noise and applying rotations to increase the amount of the training set. Secondly, we constructed a 3DCNN to extract more diverse features from multiscales. We evaluated our proposed method on the AVEC 2014. Our approach achieved MAE=7.13 and RMSE=9.16, demonstrating its effectiveness in improving the accuracy of depression assessment compared to existing approaches.

In order to meet the needs of monitoring, intelligent life, simulation training and other scenarios to study, storage, and playback of multi -source multi -made sensors, the current mainstream video codec chip has been studied, and the embedded video codec SOC FH36 is selected as the core chip as the core chip The audio and video codec and decoding platform of 16 multi-source multi-source sensors, video matrix stitching output, and H.264 coding are designed. After completing the design, fully test and verify the power consumption, delay, definition, compression rate, reliability, etc. based on FH36 audio and video codec and other aspects. Through continuous optimization, the reliability, stability, and functional performance of the audio and video codec platform have been further improved. Unilateral coding decoding 1080p@30FPS can reach 5.8Mbps, and the encoding system delay can be controlled at 1080p@30fps at 1080p@30fps can be controlled at 1080p@30fps at 1080p@30fps can be controlled at 1080p@30FPS at 1080p@30fps at 1080p@30fps can be controlled at 1080p@30fps at 1080p@30FPS. Within 100ms. Based on the embedded SOC video record storage matrix, it can achieve compressed storage of up to 16 -way sensor video information. At the same time, the video decompress stability, the screen is clear, no jitter, the system function and performance have achieved the expected goals, and mass production has been achieved.

In order to solve the problem that the existing remote sensing image change detection methods based on deep learning mostly focus on designing complex convolutional neural network models to enhance the detection capability, ignoring the counteracting elimination of background interference information of image features and the attention and characterization of target information of image features in the detection process, a VCA-Net remote sensing image change detection method incorporating variable convolution and attention mechanism is proposed based on the U-Net model, which introduces deformable convolution and attention mechanism modules in the encoding and jumping stages of the model, respectively, to adaptively acquire the feature information of the before and after changes of remote sensing images, filter and suppress the background noise interference in the detection process, and enhance the learning and characterization ability of the network model on the key feature information of the ground feature.In the experiments conducted on LEVIR-CD and CDD datasets, VCA-Net obtained 91.31% and 94.12% accuracy results, respectively, and the experimental comparison and analysis with existing model methods showed that the method has better accuracy and stability in remote sensing image building change detection.

In this paper, we propose an interactive image segmentation method based on the geodesic voting model. Basically, the segmentation is implemented by thresholding a feature map generated by a geodesic voting strategy. In contrast to the classical geodesic voting method which considers isotropic metric, we exploit state-of-the-art asymmetric quadratic metric for geodesic computation. The main contribution of this work lies at an obstacle-based geodesic computation way, in order to encourage the computed geodesic paths to pass through image features of interest. The use of asymmetric quadratic metric for computing geodesic paths can take advantages from the anisotropy and asymmetry geometric features, thus overcoming the shortcuts problem suffered in the isotropic case. Experiments on real images demonstrate that the proposed method indeed can obtain promising results.

In deep learning-based low-light image enhancement (LLIE) methods, the more effective use of image features plays a crucial role in enhancing the quality of images. In the paper, a Transformer-based multi-scale gradient feature fusion network (TMGFFN) is proposed to improve the overall image quality by making full use of the multi-scale, non-local gradient features of an image. Specifically, to exploit the non-local features of an image, an asymmetric Transformer structure is designed to enhance the image as a whole by learning the non-local mapping of the image in different directions. In addition, to extract richer features, a multi-scale structure with an asymmetric Transformer as a benchmark is designed to accomplish accurate information injection by progressive fusion of features from different sensory fields and by exploiting the contextual features of the image. Finally, a comparison with several classical LLIE methods is made. Experimental results show that our method can significantly improve the quality and sharpness of low-light images while retaining more detailed information.

Cardiac image segmentation is an important step in clinical diagnosis of heart diseases. To this day, CNNs based UNet-like networks have been applied successfully in cardiac image segmentation, yet they still fail in taking full advantage of global semantic information due to the limits of the convolution operation. On the other hand, annotation of cardiac images is a time-consuming and laborious task and requires the participation of experts. In this paper, we proposed an External Attention Augmented SwinUNet (EA-SwinUNet) network for semi-supervised cardiac image segmentation. First, we proposed a novel mean teacher semi-supervised architecture to cardiac segmentation network. Second, to overcome the limits of CNNs, we designed an improved SwinUNet segmentation network to leverage the global semantic features. By introducing external attention mechanism in skip connections, it can help network overcome the accuracy impact of small datasets and enhance the feature association between global samples. To validate our EA-SwinUNet, we conducted several experiments on the ACDC dataset and our network achieves a Dice score of 77.31% and a 95%HD of 7.03.

The next-generation of artificial intelligence technology has contributed significantly to the development of medical intelligence. However, the widespread use of deep neural networks (DNNs) has also brought about serious security threats. In this paper, we present an adversarial attack approach for deep learning-based image segmentation models in the field of medical image analysis. In our solutions, we propose a novel adversarial attack method, which is designed to exploit the DNNs’ generic down-sampling operation to ensure the effectiveness, stealthiness, and transferability of the attack. We perform the attack on two State-Of-The-Art (SOTA) models, DDANet and CaraNet in a general medical image dataset Kvasir-SEG, and a comprehensive evaluation shows that our attack is effective stealthy, and transferrable.

Aiming at the problem that the vehicle detection process in the foggy scene in automatic driving is vulnerable to the influence of fog, haze and other weather conditions, which causes a decrease in the quality of images captured by sensors, resulting in low accuracy of vehicle detection and positioning. This paper proposed a joint optimization scheme by combining DCP and YOLOV7 backbone detection networks called DCP-YOLOv7. This network consists of two modules: the image dehazing module DCP and the YOLOv7 detection module. Specifically, the DCP module performs image dehazing and YOLOv7 module performs vehicle detection. Experiments were conducted on the synthetic foggy VisDrone dataset to test the detection performance of DCP-YOLOv7. The results show that the mean Average Prediction (IoU=0.5) is 36.6% and 16.0% higher than that of SSD and Faster RCNN algorithm, reaching 86.7%, which can effectively detect vehicle object in foggy scenes.

In response to the challenge of obtaining clear monitoring images in non-ferrous metal smelters, particularly under complex atmospheric conditions and varying levels of dust and fog concentrations, we have designed an indoor dust and fog image clarification algorithm in this paper. This algorithm utilizes a concentration evaluation to obtain fog weights and construct a refined transmittance in dense fog regions. Transmittance values for non-dense fog regions are calculated using the a priori dark channel and ambient light matrix of the region. The Alpha blending technique is used to fuse ambient light and transmittance matrices of different regions, while bootstrap filtering is utilized to suppress generated noise and preserve image edge information. Finally, the global ambient light and transmittance values are obtained and substituted into the atmospheric scattering model to recover low-illumination dust and fog images. Experimental results demonstrate that the proposed algorithm achieves excellent performance in non-ferrous metal smelting environments by effectively reducing fog concentration in images, improving image illumination, and preserving image edge information.

This article proposes an improved hardware implementation method for phase-based frequency estimation algorithm. Compared with the original algorithm, this paper uses spatial Gabor filtering instead of frequency domain Gabor filtering to avoid complex spatial frequency domain conversion and reduce resource requirements for hardware implementation. The entire algorithm is designed as a pixel stream processing architecture with pipeline blocks, enabling it to easily adapt to common image sensing components with streaming pixel output, while achieving higher performance at lower system costs. This article implements a prototype based on Field Programmable Gate Array (FPGA), which can process up to 100M pixels per second and calculating vibration frequency in real-time.

In the fact of shortcomings in traditional man-machine interaction, the display console based on the intelligent voice non-contact man-machine interaction is studied in this paper, which avoids problems in incorrect manipulation and difficult operation to a extent. With the development of intelligent voice technology, speech recognition, speech synthesis and speech enhancement are widely used instead of traditional mouse, keyboard and numeric keypad. Several typical applications are introduced in the study, which can be a reference for other display consoles.

In large-span grid structures with thousands of members involved, member deformation is one of the most commonly observed damages affecting the normal service, even the safety of structures. Generally, structural maintenance measures for grid structures, such as regular test and appraisal, are required. However, Traditional manual visual inspection shows weakness in the accurate judgment of crooked members and precision of deformation measurements. Therefore, an efficient and accurate testing method is in urgent need to meet the engineering requirement. In this paper, a new deformation inspection method using three-dimensional laser scanning is proposed. The method contains two steps, firstly, the mesh model of the structure is reconstructed from 3D scans, and then crooked members are automatically recognized and their deformations are extracted by the proposed algorithm. To validate this method, an ideal model from 3D modeling and a real model from an on-site experiment are used to test the precision and robustness. Comparing the extracted and real deformations of ideal and real models, the proposed method is applicable under noise interference, and the average error of deformation extraction is less than 0.6mm for both normal and crooked members in the real model, which validates the proposed method in the on-site inspection.

Aiming at the poor application of diagnosis of infant neuromotor diseases such as cerebral palsy based on computer vison, an OpenPose-based detection model for infant cerebral palsy by extracting features from infant spontaneous motion is proposed. Firstly, the deep separable convolution and the residual network structure is used to reduce the degradation of network operation and improve the detection accuracy of joint points. Then, it gives a loss function model based on Smooth L1 to improve the detection accuracy of infant motion features. Finally, motion characteristics are assigned to support vector machine to classify the infant cerebral palsy, which realizes pre-diagnosis with several features. Experiments conducted on dataset show the accuracy of this proposed method is 7~8% higher than others and reduce the amount of calculation to 1/9 of the original and accuracy of prediction can reach 91.89%. The results show that the detection model is feasible and effective for on-line pre-diagnosis of infant cerebral palsy.

Automated pavement crack detection is of great significance to the efficiency of road maintenance. Benefit from the development of convolutional neural networks (CNNs), automatic crack detection has been gradually developed. While the convolutional neural network improves the accuracy of crack detection, the calculation amount and parameter amount of the model are greatly enhanced. This limits the application of crack detection methods on some mobile devices. In order to solve this problem, We propose a lightweight network which uses a lightweight feature extraction module combined with an attention mechanism to extract features from crack images. We use the multi-scale feature fusion module to achieve the fusion of different scale crack features. Through these modules, We built a network with nearly 0.94M parameters and only 6G FLOPs while achieving comparable crack detection performance. Extensive experiments on DeepCrack dataset show that the proposed network is superior to other comparison networks.

In simultaneous localization and map building systems, dynamic targets often have a huge impact on feature point extraction and matching tracking. This paper proposes a dynamic target detection method based on YOLOv5, which effectively improves the accuracy of feature point extraction in dynamic environments and thus avoids front-end mis-matching of dynamic feature points.

First, data gloves were made. The STM32 microcontroller was selected for the data glove to collect hand motion pose data from the MPU6050 sensor, and the data was transferred to the Unity3D engine via a wireless Bluetooth module. The experiments show that the data glove can provide data to the system stably and the system does not lag. The HMM-based dynamic gesture recognition method is proposed. The hidden Markov model is studied in depth, and the gesture interaction recognition model is established by adding noise and normalization to the data generated by the data glove. The proposed algorithm has a correct recognition rate of more than 90% for all the eight defined gestures. Finally, This paper is based on virtual reality even though the research and integration of human-computer interaction systems has made some progress and breakthroughs in the new field of virtual reality.

An improved hybrid task cascade (HTC) network for a multi-scale trash instance segmentation detection method is proposed to address the problems of ambiguous feature representation and low utilization of feature information in instance segmentation-based trash detection methods. First, multi-scale convolution is introduced in the backbone network based on the HTC network model to improve the backbone network's feature extraction capability for different sizes of trash. Second, improve the feature pyramid network structure by performing feature rescaling on feature channels based on the feature pyramid network, so that the network creates interactions across channels that do not change the spatial information, with the goal of reducing the influence of less important factors while using fewer parameters to achieve the best recognition degree. The experimental results show that the improved HTC network model can achieve robust trash feature extraction, improve detection performance, and significantly improve trash detection accuracy.

Aiming at the problems that the current brightness detection methods of airport runway edge lights is relatively traditional which has low detection accuracy and their detection speed cannot meet the requirements of technical standards, an online detection method for brightness of runway edge lights based on improved CenterNet is proposed, and key points are used to achieve classification and boundary regression. Firstly, ResNeXt is used to replace the original ResNet backbone network of the model to improve the image feature extraction ability of runway edge lights. Secondly, to address the problems that the runway edge light target is small and easy to be confused with other lights in the flight area, a feature fusion network is designed to enhance the feature expression ability of runway edge light spots. Finally, depthwise separable convolution is introduced to reduce the amount of network parameters. Experiments are carried out on the self-built runway edge lights dataset, and the results show that the proposed runway edge lights brightness detection method has strong robustness under different weather backgrounds, and the average accuracy reaches 97.09%, which increased by 2.38% and solves the problems of inaccurate positioning, multi-checks, and false detection in the original model.

Backbone networks and cross correlation operations are two important modules of the current mainstream Siamese network tracking framework. However, backbone networks based on dilated convolutions have a checkerboard effect while expanding the receptive field. The size of the depth-wise cross correlation features matching area and the number of channels activated affect the discrimination ability of the object boundary. Aiming at the problems of checkerboard effect and depth-wise cross correlation operation, proposes a Siamese tracker based on hybrid dilated convolution and global cross correlation, which uses hybrid expansion convolution to improve the checkerboard effect, improve the effectiveness of feature extraction, the global cross-correlation operation is used to promote the accuracy of the matching between the template feature and the search region feature. Evaluated on VOT2018, OTB100, UAV123 and other public test sets, the performance of this proposed method is better than that of mainstream Siamese network tracking algorithms, and the tracking effect is effectively improved.

Neural rendering can achieve satisfactory illumination effects with unknown light source locations and brightness. When training on dynamic scenarios (where object positions, textures, lighting, and viewpoints can vary), variations in object position, material properties, light intensity, and angle make the training results unsatisfactory due to the influence of global light sources. This paper proposes a novel global illumination exploration method, which uses Markov Chain Monte Carlo (MCMC) to perform small sample sampling of light points, and incorporates a new Bayesian optimization (Randomized Search) strategy to optimize the sample data combined with real-time data to remove sampling redundant information and improve training efficiency. The comprehensive experimental results show that our proposed method provides a practical and effective solution for the study of global illumination in dynamic scenes.

Pedestrian detection is widely employed in a variety of industries, including security, autonomous driving, and pedestrian flow statistics. However, because most pedestrians in real-world scenes are densely distributed with occlusion issues, scale variation and pose variability, majority pedestrian detection algorithms have low detection accuracy and a high miss detection rate. In this paper, we propose a new FPN neck module, Multiple Link Feature Pyramid Networks (MLFPN) for dense pedestrian detection, which innovative information transfer pathways and incorporation of Poly-Scale Convolution (PSconv) capable of long-range interaction into feature extraction tasks to reduce feature information loss and enhanced representation of effective pedestrian features, achieving 0.6MR^-2, 0.35AP, and 0.21JI improvements on CrowdHuman dataset, improving detection accuracy and reducing missed detection rates.

The fields of computer vision and special effects production often require editing and synthesis of facial expressions, such as transfer one person's expression to another's face. It is crucial to synthesize realistic facial expressions to ensure accurate expression of a person's feelings, however, because expressions are very subtle and subject to individual differences, facial expressions generated using methods such as 3D facial reconstruction and face morphing are difficult to guarantee sufficient facial details and the workload is huge, requiring a lot of time and effort. Facial expression transfer techniques can synthesize a large library of face expressions while ensuring human privacy, and provide data support for related face research problems. With the development of deep learning, generative adversarial networks provide a new solution for facial expression transfer. The currently available GANs are CGAN, StarGAN, G2-GAN, GC-GAN, geometric feature and attribute tagging driven GAN, Expr-GAN, TER-GAN, and StyleGAN. This paper focuses on the above methods of facial expression transfer based on generative adversarial networks, and discusses their respective features and advantages; summarizes the available facial expression transfer; finally, based on the problems of existing methods, research directions for future work are proposed.

The color assessment cabinet has important applications in the color fields of textiles, pigments, printing and dyeing, packaging, and photography. In order to achieve the spectral matching of CIE standard illuminant D65 using monochromatic LEDs, in this paper, we use genetic algorithm (GA) as the spectral matching algorithm, measure the absolute spectra of monochromatic LEDs and establish a database, optimize the parameters of the algorithm through simulation experiments, and optimize the combination of monochromatic LEDs using decreasing experiments to achieve the purpose of using fewer kinds of monochromatic LEDs to obtain a better fit with the target spectrum. The results show that the correlation indices R^2 and RMSE of 30 monochromatic LED combinations fitting D65 spectrum are 0.905 and 0.946, respectively, with high matching degree of D65 spectrum; 16 LED combinations fitting D65 spectrum are proposed by decreasing experiments, which have good fitting effect and reduce the difficulty of engineering practice.

Imaging the complete and true surface of an individual fruit is difficult. The mirror is one of the important auxiliary devices for obtaining the full-surface image. To quickly obtain the full-surface image information of the spherical fruit, this paper proposes a method for acquiring equal-scale images of the whole surface of the spherical fruit by taking one exposure using one CCD camera and multiple concave spherical mirrors, which is different from directly studying image processing algorithms. An imaging scheme configuration method is proposed, and the imaging elements are configured by using the general equation instead of the traditional trial-and-error method. The acquired image is a set of atlases consisting of a front image and a plurality of back images distributed circumferentially around the front image, and each back image is tangent to the front image. Experiments show that each image in the collection is roughly equal in size and shape, eliminating the need for image segmentation and scaling during image data processing. Moreover, this study adopts the same distortion evaluation method to evaluate the overall distortion degree of the collective image at one time, thereby improving the data processing speed and processing capacity of the machine vision system.

As a key technology in the field of computer vision, point cloud semantic segmentation has been widely used in intelligent life and has become a hot spot and a difficult area of research in 3D vision. Current research on the semantic segmentation of 3D point clouds is mainly focused on the field of deep learning, however, there are two problems with such models: insufficient local feature representation and long model training period. To this end, we propose a point cloud semantic segmentation model based on geometric segmentation and graph neural network to improve the performance of semantic segmentation and model training efficiency. In our method, Firstly geometric segmentation is used for sparse and feature representation of the original point cloud. A directed graph is then constructed to partition the point cloud into "superpoints" and "superedges" with specific feature descriptors. followed by a spatial transformer network (STN) as the backbone of embedded network to encode local features from these superpoints. Finally, the graph neural network is employed for the classification of architectural components such as walls, ceilings, floors, and beams. Experiments were conducted on the Stanford University Large Scale 3D Indoor Spatial Dataset (S3DIS) dataset. Compared with the SPG network that also takes the entire large scene as input, the proposed algorithm improves the overall accuracy (OA), mean accuracy (mAcc), and per-class intersection over union (mIoU) by 1.0%, 2.9%, and 3. 0%.

Pedestrian trajectory prediction is a key prerequisite for path planning and decision-making in autonomous driving. Most of the works based on graph neural network only deal with the dimension of spatio-temporal fusion, ignoring the trajectory features of individual pedestrians and the interaction features between pedestrians. In addition, the uncertainty of pedestrian movement and the potential inherent bias caused by the dynamic changes of surrounding environment and other factors are also rarely considered. We propose an bias correction strategy incorporating spatiotemporal attention networks to mitigate the negative impact of inherent bias on predictions in real environments. First, we extract temporal dependencies and pedestrian interactions in temporal and spatial dimensions, respectively, and construct a bias tensor to model the inherent bias; further, the bias tensor information is decoded, and then the deviation calculation is performed with the decoded original trajectory information, and then the correction of the trajectory prediction is realized. Experimental results on ETH and UCY datasets demonstrate the superiority of our strategy.

As a crucial component in the realm of computer vision, multi-object tracking has garnered widespread application in areas such as autonomous driving, smart transportation, and surveillance technology.Based on YOLOv7 and BoT-SORT algorithms, this paper followed TBD (track by detection) framework. First, YOLOv7 served as the detector tasked with identifying the target and specifying its location within the frame. Then, the BoT-SORT algorithm was used to track the target based on the target detection results. The results are then tested on the MOT20 dataset. The experimental data show that on MOT20 data set, the accuracy of multi-object tracking reaches 69.36%, and the accuracy of multi-object tracking reaches 79.11%, which effectively solves the problems of target occlusion and frequent identity switching. The experiment proves that it still has good tracking effect under dim lighting and crowded complex scenes. It can achieve better tracking for the non-rigid target of pedestrians, and has higher accuracy than YOLOv7-DeepSORT and YOLOv7-ByteTrack.

The construction of road infrastructure equipment is being vigorously developed throughout the country and an improved UNet network model is After a long period of use, the pavement problems become more and more obvious, and the emergence of neural networks has substantially reduced labour costs and improved detection efficiency. After a long period of use, the pavement problems become more and more obvious, and the emergence of neural networks has substantially reduced labour costs and improved detection efficiency. One of them is UNet, an end-to-end deep learning network with excellent recognition capability for school object detection, so it can be applied to road crack defect detection. The paper improves the coding part on the traditional UNet network structure into a residual network ResNet50 and removes the connection layer and average pooling layer, using image greyscaling The paper improves the coding part on the traditional UNet network structure into a residual network ResNet50 and removes the connection layer and average pooling layer, using image greyscaling, image denoising and image enhancement to improve the recognition accuracy. The experimental results show that the proposed algorithm improves the accuracy by 2.1% and the IOU by 5.7% compared with the original UNet network.

In space observation, star maps contain a large number of stars and noise, whose characteristics are similar to those of the targets that need to be detected. Traditional methods struggle to ensure both high efficiency and accuracy simultaneously. This paper proposes a space target detection method based on frame difference and target characteristics. Firstly, the star map is registered, then the background is suppressed using frame difference method and noise interference is reduced by Gaussian blurring. Finally, a multi-scale local target characteristic algorithm is used to calculate three feature parameters of suspected regions to further screen the targets. Experimental results show that the proposed algorithm significantly reduces false alarms while ensuring correct detection rates, and its speed is significantly improved. The algorithm fully utilizes the advantages of the frame difference method and the multi-scale local target characteristic algorithm, thus improving the detection efficiency and accuracy.

Pixel-level sea-land segmentation on high-resolution remote sensing images is a basic task in remote sensing applications and is of great significance for coastline extraction and near-shore marine target detection. This paper proposes an improved UNet model for sea-land segmentation on high-resolution remote sensing images. This model outperforms UNet++ and other models in sea-land segmentation accuracy when applied to the HRSC2016-SL dataset. Based on this model, a parallel sea-land segmentation processing algorithm was developed, whose parallel efficiency reached 49.5% on 16K*14K remote sensing images. To complete this study, a parallel processing system for sea-land segmentation was developed, which achieved distributed and single-machine multi-core parallel sea-land segmentation tasks on highresolution remote sensing images.

As one of the most serious damages of concrete structure, crack will lead to shorter service life and lower safety factors of concrete structure. The existing concrete crack segmentation model has the shortcomings of large number of parameters and slow inference speed, which makes it difficult to apply the model to the actual detection scene. To solve the above problems, this paper proposes a lightweight concrete crack segmentation method based on DeeplabV3+, which effectively reduces the number of model parameters and improves crack segmentation accuracy by replacing the backbone and adding the SSH (Single Stage Headless) context module and ECA (Efficient Channel Attention). The experimental results show that the crack segmentation Precision and IoU of the proposed model are superior to other models when comparing the unit parameter number, the crack segmentation Precision and IoU of the proposed model reach 80.80% and 71.45% respectively as well, when only accounting for 14.3% of the parameters of DeeplabV3+ with ResNet101 as the backbone.

With the increasing prosperity of shipping, effective harbor management has become a requirement. High-resolution optical remote sensing images can effectively monitor and manage ships due to their rich spatial information, clear geometric information, and unique top-down view. However, harbor scenes usually have complex background information and densely distributed ship targets, these two factors increase the difficulty of the algorithm to detect ship targets in remote sensing harbor scenes. With respect to these two issues, this paper proposes a novel harbor ship detection method based on YOLOv6-s with finer-feature fusion which is called F-YOLOv6. First, Extract ship targets from complex ground information. By importing the backbone extraction network efficient re-parameterization (Efficient-Rep) of YOLOv6-s, its efficient extraction ability can help the algorithm to improve the detection accuracy. Second, to improve the detection performance of the algorithm for densely arranged ship targets, the finer pixel aggregation network (F-PAN) is constructed to obtain more ship object position information, and on this basis, it is integrated with the deep high-level semantic information to improve the robustness of the algorithm. Third, to further improve the timeliness of the algorithm, the detection efficiency of the algorithm is improved by deleting some redundant prediction headers. Finally, extensive experiments are carried out on the harbor ship detection (HSD) dataset, which is suitable for densely arranged ship detection, to verify the effectiveness of the proposed method. In addition, the method is verified by the DIOR public dataset, and its generalization is better than other benchmark algorithms.

In order to solve the problems of poor paving effect of the existing tactile path, late maintenance is not timely, resulting in inconvenient travel for the blind, the shortcomings of the existing tactile path management are analyzed, so as to guide the design. In the design process, a defect detection method based on the YOLO-V5 target detection algorithm was adopted to realize the processing of tactile path samples, training model and classification of tactile path defect types, and a complete set of tactile path defect detection equipment was designed. This equipment can detect the tactile path in real time, judge and classify the defects of the tactile path, find out the problematic tactile path and send the geographical location information back to the background staff to assist the municipal administration department and the tactile path rectification construction unit to grasp the overall construction situation of the tactile path. Compared with the existing artificial tactile path defect detection methods, the proposed tactile path defect detection method based on OLO-V5 has better recognition performance and higher detection efficiency, and can accurately detect and identify various tactile path defects. At the same time, this study focuses on the safety of travel for the visually impaired, which helps to improve the equal rights of the visually impaired to participate in social activities and enjoy public resources, and helps to optimize barrier-free facilities, which plays a crucial role in the harmonious and orderly development of the society.

Simultaneous Localization and Mapping (SLAM) is a key technology in robot field. Aiming at the problem that traditional SLAM system has low localization accuracy in high dynamic scene, this paper proposes a visual SLAM system combining object detection in dynamic environment. Firstly, the potential moving objects in the image are located in the object detection boxes, and the feature points belong to the moving objects are filtered out. Then, the possible residual dynamic points in the environment are thoroughly filtered by combining the optical flow method. Finally, the pose of the system in the dynamic environment are estimated by using the static objects and the remaining feature points in the background. Using TUM RGB-D datasets for experiments, the results show that compared with ORB-SLAM2, the pose estimation accuracy of the improved system is improved by 90% on average under high dynamic scenes, and the positioning accuracy is also significantly improved compared with the classical dynamic visual slam system DS-SLAM.

Vehicle detection is the important basis for smart traffic. In order to further improve the performance of vehicle detection in freeway scenes, an improved YOLOv7 network is proposed in this paper. Firstly, vehicle dataset is created via cameras mounted on freeway. Secondly, Improved data augmentation method is applied to simulated the special environment interference. Then, SimAM attention mechanism module is inserted into YOLOv7, and the optimal adding position is obtained by discussing the performance of different attention composition. Finally, we provide experimental results to verify the effectiveness of the improvement. The results show that method proposed in this paper increase the mAP by 11% than original YOLOv7 on our dataset.

Depression is a common mental illness characterized by symptoms such as low mood, pessimism, and insomnia. In this study, we developed a deep Dual-Stream CNN to automatically diagnose and classify depression in expression video sequences. The network has two branches that extract static features and dynamic features from static and dynamic expressions, respectively, which are then fused for depression classification. The experiments were performed on the AVEC2014 database, and the results showed that the Dual-Stream model significantly improved the classification performance of depression, achieving an accuracy of 69.08% in depression categorization.

Accurate knowledge of soybean distribution is essential for regulating the soybean market and promoting sustainable development. This research presents a method for identifying soybean planting areas using multispectral data and feature selection models. The study area was the county of Guoyang, Anhui Province. Sentinel-1/2 data from three different time periods were used to calculate vegetation indices, textures, and other features. Initially, non-cropland elements in the study area were gradually masked by constructing decision trees. Then ReliefF algorithm was combined with Random Forest, Support Vector Machine, and BP Neural Network to develop three classification models: ReliefF-RF, ReliefFSVM, and ReliefF-BPNN, and the best feature subsets were selected as input for each of the three models, an overall distribution map of soybeans in the study area was obtained. Finally, the performance of three models was evaluated by utilizing the distribution results of soybeans obtained from the classification of planet images and seven uniformly distributed sample plots (500m × 500m) within the study area. The results showed that the Kappa coefficient was between 0.73 to 0.85, and the overall accuracy ranged from 87.42% to 93.22%. The ReliefF-RF model had the best performance in the experiment.

Aiming at the poor detection performance of UAV aerial photography based on deep learning on small targets and occlusion targets, the paper proposed an improved UAV aerial photography detection model based on YOLOv4-Tiny. By employing a lightweight YOLOv4-Tiny framework, a multi-scale detection strategy is first introduced to expand the detection scope. Moreover, the hybrid deep and shallow semantic information are utilized to enrich the feature of small and obscured targets. Second, the Efficient channel attention module is added to the enhanced feature extraction network to redistribute channel weights and the secondary screening of feature information, so as to improve target detection ability. To evaluate the proposed model, the paper used a PASCAL VOC data set and the corresponding visualization detection was carried out on the aerial photography data set. The results showed that the proposed model's average accuracy (mAP) is improved by 4.19%, and more importantly, the missing detection rate of both small targets and occlusion targets is effectively reduced. Finally, the proposed model was deployed on the Nvidia Jetson TX2 platform, and the aerial photography experiment showed that the FPS of the model on TX2 reached 30.2, meeting the real-time requirements of UAV aerial photography detection.

Timely and accurate acquisition of urban feature information and urban feature classification from high-precision 3D LiDAR point cloud data has become an international research hotspot. With the rise of deep learning, researchers have gradually considered using deep learning to deal with point cloud classification problems. However, point cloud datasets for urban scenes are different from computer vision datasets in that they have the characteristics of large amounts of data, complex scenes, and abundant category information. Applying deep learning to the point cloud classification problem of urban scenes still has great challenges, and the loss of feature information in the process of network acquisition of multiscale features is the problem that needs to be faced, and these lost features are crucial for point cloud classification of urban scenes. Therefore, from the perspective of local feature information loss, we propose EFFNet (External Feature Fusion Network), which combines end-to-end extracted features and manual descriptors using depth feature and manual descriptor technology to obtain more fine-grained local features of point clouds. Experimental results show that this method has advantages in urban point cloud classification. Therefore, from the perspective of local feature information loss, we propose EFFNet (External Feature Fusion Network), which combines end-to-end extracted features and hand-crafted descriptors using the technology of combining hand-crafted descriptors with depth features to obtain more fine-grained local features of point clouds. Experimental results show that this method has advantages in urban point cloud classification.

Real-time defect detection on industrial production lines is a goal pursued by manufacturing companies. Currently, traditional visual inspection methods are still used for defect detection on chainsaw chain production lines, which suffer from long inspection times, high missed detection rates, and high costs. In order to solve the problem of defect detection in the production process, this paper proposes a method based on an improved YOLOv5s algorithm model, and designs a new YOLOv5s-GBC model to improve detection speed and accuracy. A cost-effective convolution method is introduced to replace the ordinary convolution in the YOLOv5s backbone network, reducing the computational cost of feature extraction; Then, an additional weight is automatically calculated based on the importance of different feature layers in the Neck section, and the fusion effect of semantic information at different levels is enhanced by deleting nodes and adding input edges; Finally, a mixed attention mechanism is added to the minimum size feature output end to make the network more focused on the feature information of small targets, improving the accuracy of model regression and prediction. Through experiments, the YOLOv5s-GBC model achieves mAP of 97.4%, F1-score of 91.9%, FLOPs of 15.9G, and weight of 13.4Mb on the chainsaw chain product test dataset. The mAP and F1-score of the YOLOv5s-GBC model are 3.1% and 2.8% higher, respectively, than those of the YOLOv5s model, and FLOPs and weight are both reduced by 0.4.

In order to solve the problem that traditional fruit grading methods are easy to cause damage to fruits and low grading accuracy, taking apple grading as an example, transfer learning is used to pre-train the data-enhanced dataset, and an adaptive pooling layer is added to the feature layer of the original VGG11 model, and then an SE module based on attention mechanism is added between the adaptive pooling layer and the classifier layer to increase the weight of the effective feature map and reduce the weight of the invalid or small feature map. In order to make the training model get better results, thereby improving the accuracy of the algorithm. Comparing the improved model with GoogleNet model, LeNet-5 model, etc., the SE-VGG11 model has the best effect, the classification accuracy can reach 98.53%, and an apple grading time only takes 0.1s, which can meet the needs of factory automation classification. In summary, the improved model can be used for efficient grading of apples.

In man-machine cooperative environment, in order to ensure the safety of the cooperative robot and people, it is required that the cooperative robot can detect the collision quickly and accurately. Since the use of external sensors not only increases the complexity of cooperative robots, but also increases the cost, this paper proposes a collision detection method for cooperative robots without external sensors based on long and short term memory network (LSTM). Firstly, the dynamic trajectory is calculated by using the observation matrix of the dynamic model, and then the operating data of the cooperative robot is collected through the trajectory, and the dynamic parameters of the cooperative robot are obtained by using the least square method. Finally, the LSTM network model is designed to compensate the errors generated in the identification process of the dynamic parameters, and the results before and after the compensation are compared by experiments. The experimental results show that the recognition accuracy of the dynamic model can be effectively improved by the compensated results of the network model, and the collision detection accuracy of the cooperative robot can be effectively improved by comparing the actual torque with the theoretical torque calculated by the compensated dynamic model.

Defect detection as a part of industrial production is essential to monitor product quality. In the last decade, convolutional neural networks have been widely used in industrial defect detection, but existing CNNS-based fabric defect detection models suffer from the problems of not fully utilizing contextual information and inadequate characterization of the underlying features. To address the issues above, we propose a novel hierarchical feature fusion network with receptive field block is proposed for fabric defect detection. The low-level features are efficiently characterized by designing the context-aware feature extraction module (CAFEM) with a short connection. Then the novel receptive field block with five branches (RFB-5) can integrate different scale high-level feature maps, and the holistic attention module (HAM) is adopted to focus on the significant information. Moreover, the low-level features are fused with the high-level features to better represent the fabric texture information. Finally, a joint loss with a boundary IoU loss and a cross entropy loss is adopted to guide the network to learn more detailed information. Experimental results conducted on our built fabric image dataset demonstrate that the proposed method makes it possible to locate the defective areas in the fabric, which is superior to the seven existing methods.

Combining big data, artificial intelligence, deep learning and other technical means to allow consumer tagging to achieve accurate marketing is an intelligent new retail innovation model. To solve these problems, personalized characteristic labels for goods are established to achieve increased efficiency and reduced costs and to improve the perception of customer-consumer experience. A multi-view label prediction of personalized odour characteristics is proposed. Specifically, a gating mechanism is used to coordinate the odour knowledge graph with cigarette consumption review data. and jointly learn the weight of each data view's contribution to the multi-label features. Then, combining the prediction results of all classifiers and the learned contribution weights, a final prediction can be made. The research is experimentally validated as pioneering work in the field of odour tagging recommendation and also has good effectiveness.

Wheat seed appearance quality detection is a preliminary step to obtain high-throughput phenotypic and wheat breeding information. However, the low portability due to parameter redundancy is a challenge for general detection methods based on deep learning. To address this issue, we designed a lightweight model (Nanodet-Ghost) for this task. The feature extraction structure of GhostNet optimized by ELU is applied to extract the features of wheat grains. Shallow and deep features are integrated to improve the discriminability of wheat grains. DIoU loss was used to improve detection accuracy and convergence speed. 6414 images of healthy, moldy, damaged and sprouted wheat grains were collected and divided into training set, validation set and test set in the ratio of 4:1:1. After 300 times of network training, the overall results showed that Nanodet-Ghost is suitable for quality detection of wheat grain appearance.

To improve the detection performance of overlapping apples, as well as apples occluded by tree branches and leaves, we propose an improved YOLOv7-based method for apple target detection on the MinneApple dataset. Firstly, we introduce the RepLKNet module in the backbone network, which directly enlarges the kernel size to obtain a larger effective receptive field (ERF) and extract more feature information from the images. Secondly, we apply the Wise-IoU loss function and a dynamic non-monotonic Focusing Mechanism strategy to adapt high-quality anchor boxes to low-quality samples, and pay more attention to normal anchor boxes, thereby improving the overall performance of the network model. Compared with current mainstream algorithms, our proposed method has certain advantages, achieving a Mean Average Precision (mAP) of 96.17% and a detection efficiency of 39 frames per second (FPS) on the MinneApple dataset, meeting the real-time detection requirements.

Facial expression is a basic way to express human emotions, and it is the primary medium for individuals to communicate with others. The display of facial expressions can better understand the other person's feelings and emotions. With the continuous development of artificial intelligence technology, the demand for human-computer interaction has also increased. The development of artificial intelligence plays an important role in face recognition and micro-expression changes when users watch IPTV. Because of the small range of motion and fast change of micro-expression, it is difficult to analyze it manually. Therefore, it is very necessary to develop a reliable recognition system. In this paper, the face recognition model facenet model is used to learn the image features of the face, Yolov5 and Yolox neural network are used as the basis to compare the models trained, and the performance of the selected model is tested to select the best effect.

The relationship between modern education development and artificial intelligence is getting closer and closer, but the technology of online examination and test detection needs to be improved. Most of the existing technologies are aimed at the comparison of pictures per frame, head detection, limb detection, key part detection, etc., although this effect is ideal, but at the same time the requirements for conditions are high. In this paper, for online exam detection, it is proposed to use face recognition and Euclidean distance combined with Yolov3 to complete the cheating detection.

RGB-T tracking aims to capture both position and scale of a specific target, with the guidance of both visible and thermal images. With the popularity of multi-modal sensors, this topic has reached more and more attention, which has great potential on autonomous driving, smart monitoring, etc. Recent methods mainly introduce feature fusion modules to aggregate multi-modality information via feature selection or feature fusion. In this paper, we design an adaptive multi-modal decision fusion strategy for Visible and Thermal (RGB-T) tracking. First, we set correlation filter tracker as our baseline. Then, we calculate the tracking confidence of both modalities via Peak-to-Side Ratio, generating the fusion weights. Finally, the response maps are linearly summed via the fusion weights. Experiments on GTOT validate the proposed fusion strategy can provide superior performance against the competitors, achieving 25.3% MSR and 43.9% MPR.

Structural health monitoring is an important way to ensure the safe and reliable work of the structures during the service period. Surface crack is one of the important indicators for monitoring and assessing the damage of structures. However, crack detection still mainly relies on human visual observation and recording. Besides, people could only record the crack width of a few points instead of all points. To solve this problem, this paper aims to propose a method which can automatically detect crack and calculate the width of every point on the crack but not several points. Thus, this paper built a large database containing more than 100,000 pictures which were taken in different scenarios, and used this database and self-designed deep convolutional neural network to detect cracks. And the prediction accuracy of CNN reached 95%. In addition, combined with image processing technology, this paper further carried out crack extraction and width calculation. Thus, the whole process of crack detection, extraction and width calculation is fully automated, and the calculation results were proved to be close to the actual results, which meets the engineering requirements and perform better than traditional OTSU algorithm.

To address the problems of uneven feature points extracted by traditional SURF algorithm in image stitching, low correct matching rate and high time complexity, this paper proposes a fast image stitching algorithm based on improved SURF features. Firstly, we use the advantages of robustness and low complexity of SURF algorithm to detect feature points; secondly, we use the advantages of FLANN algorithm to adjust parameters for accuracy improvement to match feature points; thirdly, we use RANSAC algorithm to remove feature points with incorrect matching and perform the calculation of single response matrix to find the best model matching pair and reduce the error; finally, we use OpenCV Stitcher is used for image stitching and fusion. The experiments show that the algorithm has better robustness than the traditional SURF algorithm under various complex situations such as image angle rotation, brightness change, scale scaling, etc., and can better improve improve the matching accuracy and matching efficiency.

Aiming at the problems of slow segmentation speed and excessive computational complexity in practical engineering applications when using existing models to segment glass transparent containers, an efficient image semantic segmentation algorithm based on improved DeepLabV3+for glass transparent containers is proposed. The proposed algorithm uses the MobileNetV3 network to replace the backbone feature extraction network Xception of the original model, effectively reduces the number of parameters of the semantic segmentation model, improves the ASPP module, introduces the strip pooling module (SPM) and depthwise separable atrous convolution (DCAC), and uses the densely connected multiple receptive field cascade (MRFC) to obtain multi-scale feature information. In the decoding structure, multi-level feature fusion (MLFF) is used to recover the details and levels of features lost during the down sampling process. Adding the spatial attention mechanism module SAM helps refine the target boundaries. The proposed model reasoning speed is 1.5 times faster than DeepLabV3+, and FLOPs are reduced by 71.4%. In addition to a 0.14% decrease in the cuvette's M_iou, the beaker and reagent bottle's M_iou increased by 0.19% and 0.18%, respectively. From the experimental results, we have improved the reasoning speed without completely losing accuracy, achieved the segmentation effect that is comparable to the original model, and well considered the accuracy, computational complexity, and reasoning speed, achieving a good balance between various aspects.

Simultaneous Localization and Mapping (SLAM) plays a critical role in the field of intelligent mobile robots. However, the traditional visual SLAM framework is based on the strong assumption of a static environment, which is not applicable in dynamic real-world environments. This paper presents a SLAM algorithm that combines semantic and depth information. Firstly, we combine object detection and geometric methods to identify dynamic regions. Secondly, dynamic and static feature points are distinguished using a depth information classification method in these dynamic regions. Then, we eliminate the dynamic feature points and only pass the static feature points to other threads. Finally, we evaluate our approach on the public TUM RGB-D datasets and compare it against other state-of-the-art methods. The experimental results demonstrate the proposed SLAM algorithm has good accuracy and stability in dynamic environments.

Image matching is an essential technique in computer vision for many applications (e.g., image understanding). In order to improve the adaptability of previous matching algorithms to multi-source images (i.e., perspective and panoramic images) and increase the matching accuracy, we propose an improve Oriented FAST and Rotated BRIEF (ORB) matching algorithm. We first use the K-Nearest Neighbor (KNN) algorithm to roughly match the feature points extracted from the uniformly partitioned image grids and calculate their matching scores to obtain high scoring matching pairs. Then, we utilize the regional local consistency constraint and affine transformation verification of the Adaptive Local Affine Matching (AdaLAM) method to further refine the matching pairs. Finally, we perform matching experiments on both perspective and panoramic images and show better matching results than most previous matching approaches.

With the rapid development of deep learning, object detection methods have made significant progress, but small object detection is still a very challenging task. Based on object detector Single Shot Multibox Detector (SSD), we conjecture this is due to two factors: (1) In Convolutional neural networks (CNN), shallow feature maps contain strong spatial position information and insufficient semantic information. Therefore, single-layer feature detectors could not make full use of the limited information of small objects in the images; (2) The loss is not a good measure of the relationship between the prediction box and ground truth, what’s more, it is easy to cause the unbalanced loss attribution, thus impairing the accuracy of small object detection. To solve the above problems, we propose an improved SSD algorithm, named PASSD, including a multi-scale feature fusion module, which provides contextual semantic information and enhances high-resolution spatial information for shallow layers. Furthermore, a feedback-based loss function based on Gaussian Wasserstein Distance is introduced to measure the similarity between boxes in a new way. Experiment results show that the proposed method outperforms the previous method, achieving good speed and comparable accuracy trade-off. For 300×300 input, the proposed method achieves 86.28% mAP on the PASCAL VOC dataset, respectively. Especially APs is improved by 4% on MS COCO dataset compared to the benchmark model SSD.

Underwater binocular detection is a detection method using binocular stereo vision. Binocular stereo vision is increasingly used in robot vision, aerial mapping, reverse engineering, military applications, medical imaging and industrial detection. The AUV (Autonomous Underwate Vehicle) equipped with binocular optical system restores the real underwater three-dimensional world through binocular stereo detection results, and becomes an important means for underwater resource detection and underwater target search. Due to the imaging characteristics of underwater images, there are a large number of weak texture and repeated texture areas, and the current underwater stereo matching algorithm has poor matching effect in these areas. To solve this problem, this paper proposes a multi-cost fusion BP(belief propagation) algorithm based on image segmentation information to improve the matching effect of such regions. Experimental results of Middlebury data and real underwater binocular image data show that the proposed algorithm improves the matching effect of weak texture and non-texture regions.

It is necessary to develop an efficient method for the detection of seedlings shortage and breaking rows of soybean. In this paper, we design an efficient detection algorithm for the shortage of soybean seedlings using RGB images captured by UAVs. Then we apply a series of techniques for detecting the centerline of the crop row. Otsu-based binarization segmentation is utilized to segment the vegetation canopy from the image, thereby supporting the further analysis of crucial field indicators. Finally, we perform local projection segmentation of crop rows using the projection integral curve segmentation algorithm, which can detect the position of the missing seedlings in the crop rows. The results show that the accuracy of crop rows detection achieves 93.9%; the accuracy of crop rows detection under different planting modes and the UAV flight heights achieves more than 92%, which indicates the robustness of the detection algorithm. The detection results of this method have obtained a good performance inaccuracy. From the average time of crop rows detection is 0.31s, the detection speed and efficiency are also guaranteed. The analysis shows that the X coordinate of variance detection of line breaking position coordinate is P=0.97125, and the Y is P=0.13156. There is reason to believe that the test results are good.

The emergence of a new coronavirus pneumonia in 2019 has had a major impact on the world, this virus can be transmitted through airborne droplets and other carriers, and wearing a mask correctly in public can effectively prevent the transmission of the virus. Therefore, this paper proposes a lightweight face detection algorithm for wearing masks that can be applied to embedded devices. The algorithm introduces a contextual attention module based on the PP-PicoDet algorithm, which enhances the connection between feature contexts, makes the model pay more attention to face mask wearing-related features, and uses migration learning to help the model learn face mask wearing features better. A dataset of 8967 images, including both normal faces and mask-wearing faces, is created and manually labeled for network training. The experimental results show that the proposed method can effectively perform fast face-mask wearing detection, and can run as fast as 59 frames/second on the NVIDIA Jeston Xavier NX embedded device.

This paper introduces the principle and structure of the film digitizing system,and through the scanning and analysis of nuclear film, it proves that after the traditional film A/D signal conversion,the digital film meets the DS level of ISO14096 in terms of the sensitivity of the image quality meter,defects and detail display.In addition,the digital technology of film is studied from three aspects: negative signal transformation mode,superposition and splicing processing function and defect evaluation. Firstly,the digital software is used to superposition and splice the original signal of the film to evaluate the double film, so as to reduce the defect detection rate of the overlapped part. AI intelligent system can helpautomatically identify defects, reduce the risk that the detection results are easily affected by human factors, and is better than traditional ray defect detection which mainly relies on manual evaluation. In particular, the 3D modeling module can extract and analyze the negative data signals to complete the assessment of the negative quality, which has its unique advantages and great significance for the optimization of welding and manufacturing processes.

To address the problem that FPN in the one-stage object detection algorithm loses a large amount of semantic information after upsampling and downsampling fusion, leading to difficulties in detecting small objects in dense scenes. This paper proposes a new FPN structure, Residual dilated FPN (RdFPN), compared to the traditional FPN, which uses multiple Dilated Block instead of the standard convolution in the FPN horizontal structure to mix the receptive fields of different layers in the feature map and expand the object size range of the feature map, while improving the network's ability to extract features at different scales. Experimental results show that RdFPN improves the mAP of the one-stage object detection algorithms RetinaNet, FCOS, and GFL by 0.9%, 0.7%, and 0.8% on the MS COCO dataset, enough to improve the overall performance of the detector.

Proposed a method for calculating the area of road diseases using a monocular camera. The camera and the road plane are calibrated to obtain the mapping relationship between pixel points and spatial points; the Mask R-CNN is used to detect and segment the road disease; the inverse perspective transformation is used to convert the target image into a bird’s-eye view to obtain the relationship between the pixel scale and the real scale; binarize the bird’s-eye view according to the segmentation effect, and the area of the road disease could be obtained through calculating the number of target pixels. It has been verified that the calculation error of the algorithm is about 2.1675% in the near vicinity.

Since the tremendous advancement of face recognition technology, how to detect real human faces from images has gradually captured the attention of researchers. To address the problem that most face anti-spoofing approaches ignore context information and local information about faces, this paper proposes a face anti-spoofing algorithm that fuses multi-scale features and texture features of images. On the one hand, we extract multi-scale features using adaptively spatial feature fusion (ASFF) and polarized self-attention (PSA) mechanism. On the other hand, local binary patterns (LBP) is used to form an auxiliary self-supervision branch to extract image texture difference information. As a final step, end-to-end live detection is achieved by balancing the difference between two different loss functions based on homoscedastic uncertainty. The EER of the presented method is 0.81% and 1.26% on CASIA-FASD and MSU-MFSD, respectively, and the ACER is only 1.64% and 2.07%±1.05% on the second and third protocols of Oulu_NPU, which indicates the validity of the method.

Machine learning (ML) graph convolution network (GCN) has been applied in various fields, including recommendation systems, and has achieved great success. The core goal of the recommendation system is to infer user preferences so as to actively provide users with valuable and targeted information. Collaborative Filtering (CF) is a classic and widely used recommendation method. However, GCN-based recommendation methods still have many problems. For example, GCN recursively merges messages from different order neighborhoods, it is difficult to distinguish between different node messages, which leads to training difficulties, and problems such as over-smoothing have great constraints on the recommendation model. Different from the current GCN-based method, in view of the above problems, it is proposed to use a simple GCN model to separately aggregate neighbor messages in different orders for collaborative filtering, and then aggregate them in a hierarchical manner without introducing other model parameters. After that, the dropout idea was transferred to the model, by generating multiple views of the same node from the three dimensions of nodes, edges and graphs to reduce the effect of over-smoothing, which prevented over-fitting and improved the performance of the model. The experimental results on three datasets demonstrate the effectiveness of the proposed model.

In order to better segment the sandstone particles, this paper improves the traditional image acquisition method by adding a baffle plate and a divider plate to collect the cohesive sandstone images. The collected images are processed by a series of OpenCV libraries in the computer, including grayscale and grayscale equalization, filtering noise reduction, binarization, morphological optimization and image segmentation based on the improved watershed algorithm. In the sandstone particle segmentation experiment, By comparing the results obtained by the image method with the results obtained by the manual screening method, the experimental results show that the image method proposed in this paper can effectively segment the real sandstone particles in the sandstone segmentation of the mixing production line, and the availability of this method is very high, meeting the industrial requirements.

In recent years, Kernel Correlation Filter (KCF) has been a highlight topic due to its excellent performance on visual tracking tasks in real-scenes. However, fixed search boxes which are used to delineate candidate objects are widely used in the current KCF based algorithms, and its tracking ability degrades when the scale of the object changes. Besides, occlusion always make the tracked object lost in real scene as the tracker highly relies on the response maximum. To improve robustness and reliability, we propose a scale-adaptive target tracking algorithm based on KCF to meet the challenge of the scale variation. At the same time, we propose a new strategy for template update with the average peak correlation energy to reduce the drift problem. When both the average peak correlation energy and the maximum response peak of the current frame are lower than the certain thresholds, the proposed re-detection model will be activated to detect the corresponding object through the particle resampling strategy and regain the position. Experiments on public data sets OTB2013 are used to verify the behaviors of the proposed method, and the results indicate that the proposed algorithm holds the superiority to the current correlation filtering algorithm in the evaluation index of success rate and range accuracy, and is more stable in the case of large scale variation of target and occlusion.

In recent years, the fashion industry has developed rapidly, and the image is one of the main carriers for people to obtain clothing information. Therefore, the analysis and processing of clothing image through the technology in the field of artificial intelligence has become a major research content in the field of clothing development. Image segmentation technology can be used to segment different clothing regions from the clothing image in order to facilitate the subsequent analysis and processing of clothing image, which is one of the basic research directions of clothing image analysis. At present, most of the methods belong to the traditional segmentation methods, or are based on the deep convolutional neural network (DCNN). Especially for Deeplab, FCN and their improved network segmentation models, although these network models have different degrees of improvement compared with the traditional methods, there are still some problems such as the fitting degree of clothing region segmentation is not high and the effect of clothing edge segmentation is poor. To solve these problems, this paper adds object context information extraction module and edge optimization module based on Deeplab v3+ network. The object context information extraction module focuses on the context information aggregation method in semantic segmentation, and enhances the feature representation of pixel points with the help of the feature information of the corresponding clothing region, so as to improve the label classification effect of pixel points in the clothing image and improve the fitting degree of clothing region segmentation. The edge optimization module makes the pixels in the edge area directly use the label prediction of the pixels in the same category area to greatly improve the edge segmentation effect. Our method performs well on the DeepFashion-MultiModal dataset, and experiments show that the improved Deeplabv3 + network outperforms the original Deeplabv3 + network and other popular segmentation methods, with an excellent mIOU of 67. 29%.

Aiming at the problems in the actual mining process of the Apriori algorithm: it does not take into account the relationship between the influence of the item, and cannot mine the association rules of some small probability events, frequently scan the database and generate a large number of candidate itemsets, an improved Apriori algorithm based on the influence weight is proposed. Firstly, according to the impact of items on affairs and the relationship between items, a weight distribution method based on item influence is used to weight items, so as to extract more hidden and valuable associated information. Secondly, in order to reduce the scanning times of the database and the overhead of I/O ports, the weight vector and compression matrix are introduced to represent the transaction database. Finally, it is compared with the classical Apriori algorithm and the Apriori algorithm based on compressed matrix. Experimental results show that the improved Apriori algorithm has higher efficiency, and can mine out frequent patterns with high influence.

The Poisson multi-Bernoulli mixture (PMBM) filter is an effective tracking framework for tracking multiple extended objects. However, methods based on this framework typically assume that the object's shape is an ellipse, which cannot adequately describe the object's shape. When two objects are spatially close, problems such as difficulty correctly distinguishing the object trajectory and insufficient utilization of the object shape's feature information arise. To address the aforementioned issues, this paper describes the shape of the object in greater detail using the multi-ellipse model. On this premise, the extended object shape will be divided into multiple Gaussian inverse Wishart components, and the likelihood will be calculated. Furthermore, a new partitioning method is proposed to divide the measurements into several most linearly dependent components, which aids in estimating the object shape composed of multiple ellipses. The simulation results show that the new algorithm outperforms the original PMBM algorithm in terms of accuracy.

In response to the issues of data sparsity, limited scalability, and low recommendation accuracy faced by traditional collaborative filtering algorithms used for personalized user recommendations in the face of increasing data volumes, an optimized collaborative filtering algorithm is proposed that addresses these problems by improving the initial clustering center and incorporating Weighted Slop One filling. This algorithm first uses user similarity as the basis for clustering, and utilizes the Weighted Slope One algorithm to predict the user-rating matrix. Then, the generated matrix is clustered using the AP-MMD-K-means algorithm to obtain a set of similar users. Finally, nearest neighbor search is performed on the target user within their corresponding similar set to generate the recommendation results. Comparative experimental results show that this algorithm improved both recommendation performance and accuracy compared to traditional recommendation algorithms. This algorithm effectively alleviates the problem of data sparsity and improves scalability, significantly improving recommendation accuracy.

Currently, the majority of multi-population algorithms ignore the worst particles and instead learn from the gbest or lbest. Hence, it is simple for these strategies to result in inadequate population variety, which causes early convergence. A multi-population particle swarm optimization considering the worst particles into account (CWMPSO) is suggested to demonstrate the importance of the worst particle information. In addition to considering the best particle, the CWMPSO algorithm concentrates on the worst particle globally. The algorithm divides the population into three subgroups, which learns from the gbest, gworst, and gbest + gworst, respectively. The development of the global worst particle information can be achieved by group learning, allowing for a more thorough exploration of potential optimum solutions and an avoidance of local functional ones. 30 complex CEC2017 benchmark functions were used in this work to test a range of algorithms including CWMPSO with various dimensions. The outcomes demonstrate that, in every dimension, the CWMPSO algorithm is better than other advanced PSOs.

With the continuous development of science and technology in China, the popularity of computers has brought a great impact on people's production and life. And all walks of life have also been continuously studying computer vision technology, and in continuous practice, it has been applied to multiple fields, such as industrial and agricultural production, human-computer interaction, social public security, virtual reality, and so on, making people's lives enter a new form. This study mainly analyzes the advantages of the application of computer vision based human behavior recognition and specific human behavior recognition methods.

In the era of rapid development of deep learning, the popular transformer in recent years has played an outstanding role in many fields, but most of it focuses on the development of language models, and is still in the development stage in computer vision. Gesture recognition is a relatively popular module in computer vision, but the mainstream method still uses CNN for gesture recognition, and there are few articles combining gesture recognition with attention mechanism and transformer. In this paper, we propose a dynamic gesture recognition model based on attention mechanism and Transformer. In order to extract the valid information in each frame, we add the attention mechanism to the feature extraction network, followed by passing it into the transformer to predict the hand gesture through the self-attention mechanism. The model has been tested on the IsoGD dataset, achiving good experimental results. Not only do we confirm that the attention mechanism can improve the recognition accuracy through ablation study, but also prove that transformer is feasible to process and identify the temporal information among frames in dynamic gestures and perform recognition.

From the vehicle guarantee big data platform, the vehicle listed data, maintenance data and use data are extracted, through data cleaning, data statute and data transformation, build the indicator system, using the improved K-prototype clustering algorithm, the vehicle use unit is divided into key control unit, key tracking unit, key concern unit, general tracking unit and general concern unit, and by analyzing the different characteristics of each group and its distribution area, it provides a theoretical basis for taking targeted management measures and reducing the proportion of faults, and provides auxiliary decision-making support for vehicle security.

To achieve non-contact automatic measurement of packaging cases in the transportation process of the logistics industry, such as secondary packaging, transshipment, and warehousing, a machine vision dimension measurement system based on binocular stereo vision was proposed. Firstly, the method of combining the SURF algorithm with FREAK descriptor improves the execution efficiency of the algorithm while ensuring the acquisition of better identification feature points; Then, the MISMATCHING feature points are eliminated through the PROSAC mechanism, and the initial disparity map is optimized using bilateral filtering. The spatial information of the measured object is obtained through three-dimensional reconstruction; Finally, a combination of Canny edge detection and Harris corner detection is used to extract the edge information of the object, and the geometric description of the external contour of the target object is obtained through line fitting and calculate the precise size of the measured object through the modified three-dimensional space volume. The experimental results show that this method can effectively identify and describe the shape of various specifications of the measured object, and the measurement error and real-time performance can solve the industrial requirements of the logistics industry.

The collected seismic data are highly contaminated due to random noise such as ambient noise, high and low frequency noise and instrument self-noise, which in turn affects the subsequent processing process. With the success of deep learning in image noise reduction, the application of deep learning to noise reduction of seismic data has also obtained better results than traditional noise processing methods. In this paper, we propose a noise reduction method for seismic data based on subspace projection and fast Fourier convolution (FFC). The method uses fast Fourier convolution to divide the channel into two branches, local and global, after downsampling, to update the feature map from two directions, and generates subspace vectors to separate seismic signal and noise according to the seismic data features to construct a seismic data noise reduction network. Experimental results show that the proposed method outperforms traditional seismic data noise reduction methods, such as wavelet transform, Fourier transform noise reduction, and classical deep learning methods FFDNet and NBNet, in terms of signal-to-noise ratio (SNR) measures.

With the continuous development of society and economy, people have more preferences and pursuits for various kinds of art works, but few people can reproduce the great works of artists like Van Gogh, which is undoubtedly a pity. Related applications show that the image style transfer technology can transfer the styles of various art paintings to common images, so that different art styles can be generated relatively quickly. Combining with the current mainstream migration ideas, this paper proposes the decoupling and extraction of style features using effective channel attention (ECA) module, the decoupling and extraction of content features using spatial attention (SA)module, and the feature fusion using the feature fusion (FF) module. The proposed method is compared with the mainstream method in terms of speed. The results show that the method used in this paper has better results in both speed and sensory. The proposed method can satisfy people's pursuit of artistic works, inspire contemporary artists to create, and promote the cross integration of art field and computer vision field.

Efficient and accurate segmentation of the basic features in the image of sugar orange is the primary work for intelligent classification of sugar orange. In order to achieve the purpose of grading according to the key quality characteristics of sugar orange, the peel region, fruit stalk region and peduncle region were segmented respectively, and the basic feature information of each region was extracted. Firstly, the background of the feature image was segmented, and then the fruit diameter, shape, ripeness, texture, color distribution and other features were extracted for grading.

To address the problems that the chimpanzee optimization algorithm is prone to fall into local optimum and convergence speed is too slow. This paper proposes a multi-strategy improvement of the chimpanzee optimization algorithm by fusing differential variation and random variation (DRChOA). Firstly, an adaptive weighting factor is introduced to balance the global search and local exploitation ability of the algorithm, secondly, the differential variation strategy is introduced to facilitate the exchange of information among individuals for population position updating, and finally, the random variation strategy is used to enhance the ability of the algorithm to jump out of the local optimum.

Asphalt pavement detection is a problem characterized by low efficiency, high manual input, and high cost. A 3D visualization management platform for road infrastructures based on artificial intelligence and information technology is presented in this study. Initially, images of cracks and potholes were processed digitally and segmented using image processing methods. A pavement condition index (PCI) is calculated by combining pavement distress identification with cracks and potholes. After that, ground penetrating radar (GPR) was used to detect the road structures, and then Canny edge algorithm was performed to process the original GPR images. Cracks, looseness, and poor interlayering were identified with the object detection model and data enhancement methods. Inverse time migration (3D) imaging of the internal diseases was performed in three dimensions (3D). Based on the results of the object detection model, F1 score and mAP were both above 80% for road distress detection. As a final step, a BIM model of asphalt pavement was generated using Revit software, and 3D asphalt pavement model rendering and quality details were displayed in Web-BIM browser, suggesting that this method can be used to maintain highways in a fast, intelligent, and scientific manner.

With the popularity of vehicles, traffic safety issues are becoming more and more important. In the field of assisted driving and intelligent driving, traffic sign detection is particularly important, which can provide adequate response time for drivers and indirectly protect the safety of people's lives and property. At present, although the development of traffic sign detection algorithms has been relatively mature, there is still some room for improvement. For example, the detection accuracy of traffic signs is not high enough, which is easy to cause false detection and missing detection of signs. Therefore, this paper proposes a traffic sign detection algorithm based on deep learning, which is based on yolov5 model and embedded with SE attention mechanism. transformer module is introduced into the backbone network framework to improve the feature extraction capability of the network. Finally, the weighted BIFPN module is used in the feature fusion network. Experiments show that the improved algorithm proposed in this paper has a good performance in the Chinese traffic sign data set, with the average accuracy of multiple categories increased by 3.6%. The detection accuracy has been greatly improved.

With the rapid development and maturity of the smart industry, smart home security systems are becoming popular. However, problems such as poor optimization under face recognition integration and high latency of smart home gesture recognition control still exist. In this paper, in response to the problems raised above, the more mature CNN algorithm and OT prime tracking algorithm are used for face recognition, and the Euclidean distance is introduced as the adjustable threshold for face recognition, and its recognition response time is 29-30ms; in gesture recognition, the mathematical model of finger angle recognition is used for gesture recognition, based on the Mediapipe open source framework, and in the experiment, its recognition time tends to be 60-100 frames, with an average response time of 61.75ms. These advances in facial and gesture recognition in home security systems provide faster and more accurate recognition, thereby enhancing the owner 's safety measures. The results of this study provide relevant insights into the development of smart home security systems and their implementation in real-world scenarios.

This paper proposes the application of deep learning-based semantic segmentation in Unity scene construction as a means to enhance efficiency and quality. By automatically recognizing and segmenting various objects in an image, this method enables developers to rapidly generate 3D models and textures for the scene. The study showcases the effectiveness of the proposed technique in reducing the time and effort required for scene construction through experiments conducted within Unity.

Underwater acoustic target recognition has been faced with significant challenges due to the noise in the ocean environment and the complex and ever-changing nature of ocean channels. this paper proposes an underwater acoustic target recognition method based on a convolutional neural network and multi-feature fusion. Various features including the Amplitude Modulation Spectrogram, Mel Frequency Cepstral Coefficient, Relative Spectral Transform-Perceptual Linear Prediction, Gammatone Frequency Cepstral Coefficient and Delta feature of underwater acoustic targets are effectively extracted and then fused to form AMCG-Delta features. To address the issue of data scarcity, data augmentation techniques including pitch shifting, time stretching, random addition of noise and SpecAugment are used. Finally, the Ecapa-OLS network is proposed to improve the accuracy of underwater acoustic target recognition. With the shipsEar dataset, the proposed method achieves a recognition accuracy that is 6.95% higher than that of the baseline method.

Human pose estimation is attracting increasing attention from researchers as a fundamental technology for human action recognition, while edge computing applications pose higher challenges. In this paper, we propose an optimization human pose estimation method for edge computing applications. This method uses YOLOv4 to improve the accuracy of AlphaPose at the object detection stage and accelerates human target detection and pose skeleton node inference by optimizing both the object detection model and the pose estimation model. Moreover, the proposed scheme is transplanted onto an embedded development board and compared with current mainstream human pose estimation algorithms. Experimental results show that the proposed algorithm achieves better detection frame rate and accuracy than the compared algorithms on the embedded platform. The algorithm has high real-time performance and accuracy and provides a basis for lightweight human action recognition scheme in edge applications.

This presentation, originally published on 1 August 2023, was withdrawn on 7 August 2023 per author request.

In this paper, the nondestructive testing and defect image signal recognition system for special welding of steel structure is researched and developed. The system not only makes use of the powerful computing power of the computer, integrates the relative software and hardware resources, but also realizes the functions of defect location, identification, display, channel control, storage and printing, etc., which greatly improves the cost performance of the instrument, and at the same time improves the intelligence of the instrument, making manual flaw detection more simple and flexible. The combination of modern digital signal processing and pattern recognition technology promotes the welding nondestructive testing technology, which not only improves the accuracy of testing, but also obtains the quantitative results of testing, such as the size, location and nature of welding defects.

In the transmission and transformation network, line status inspection is always a key, important work items. The key object of line inspection work is mainly line status and insulator status, etc. For line inspection work with security risks and low efficiency, the insulator can be used for image acquisition by UAV and image analysis by computer. The goal of image segmentation and insulator alignment is achieved by using binarization segmentation and image transformation on the captured images from the UAV. Various operations are performed based on the geometric features and pixel distribution features of insulators, and fault evaluation and feedback are realized. The results show that the designed insulator damage intelligent discrimination technology can accurately locate insulators and identify defects.

Electronic non-stop toll collection system (ETC) is the most popular highway toll collection system at home and abroad. Based on ETC gantry license plate data, this paper systematically describes the highway toll collection process and points out that abnormal data in the license plate data will significantly reduce the accuracy of subsequent data fusion and toll collection. And summarizes the abnormal situation of ETC gantry license plate data. Aiming at the spatial-temporal characteristics of such data, the ETC gantry license plate data abnormality recognition method based on an improved DBSCAN is proposed to realize the adaptive clustering separation of data, which could extract the average travel speed feature set and traffic volume feature set to recognize the abnormal data. The results show that the recognition rates of average speed and traffic volume reaches 96.47% and 88.47%, and the false recognize rate decreases to 0.71% and 0.91%, which could consider data spatial-temporal correlation for abnormal data recognition and solve the problems existing on the current application of ETC gantry data.

The vehicle object detection and tracking on drone-captured scenarios will play a significant role in the intelligent transportation in future. In order to provide accurate orientation and scale information, the angle vector module is introduced to improve the model to realize the rotated detection. For the small and densely distributed objects, the C3_TR structure is proposed to improve the detection accuracy. The image weight and the category weight are used in training to improve accuracy on the self-built vehicle dataset with the category imbalance. Finally, the mAP of the improved YOLOv5 detector is 82.4% higher than the baseline model by 8.1 percentage points. The angle parameter is taken into account and the rotated IOU matching is used instead of the horizontal IOU matching to realize the tracking with rotated bounding boxes by DeepSORT. In tracking evaluation results on the self-built dataset, the HOTA, MOTA, MOTP and ID_SW are 76.86%, 92.8%, 81.3% and 2, respectively.

Aiming at the problems of slow image processing and detection speed of flame recognition method and sensor measurement vulnerable to light environment interference, this paper studies how to improve the speed of flame recognition and overcome light environment interference. To improve the speed of flame recognition, this paper selects convolutional neural network technology, trains a classifier dedicated to flame, improves the preprocessing and postprocessing methods of feature map, and develops a new flame recognition model, which greatly improves the speed of flame recognition; To overcome the interference of light environment, this paper proposes a new method by measuring the two output signals of the flame sensor in different light environments. It provides a new method for fast flame recognition of fire-fighting robot.

Face anti-spoofing(FAS) is an important task to secure face recognition systems from presentation attacks (PAs). Most of the FAS methods use scalar(0,1) or handcrafted pixel-wise labels(binary mask) as the ground truth labels. However, these labels may are not the most adequate way to supervise PA detectors learning intrinsic and discriminative spoofing cues. In this work, we present an adaptive and learnable label for more efficient supervision of PA detectors. Then we propose a dual-branch network for FAS that utilizes the proposed supervision to detect PAs more reasonably. We evaluate our method by conducting experiments on public datasets, the proposed supervision method can not only improve the detection performance of the PA detectors, but also enhance the model’s interpretability (i.e., locating the pixel-level positions of PAs more reasonably).

Recognition of handwritten mathematical expressions (HMER) is a branch of Optical Character Recognition (OCR). In this paper, a seq2seq model with encoder-decoder structure is constructed based on ResNet and Transformer deep neural networks, and end-to-end training is performed using open source datasets. The results of the experiments name that the character error rate of the model proposed in this paper is improved by 11.36% compared with the CNN and LSTM-based HMER model.

In this paper, we propose a new method for driving style recognition on the basis of SAX and bitmap technology. This method converts the sensor signal into a discrete symbol sequence and uses a bitmap for feature extraction. The weight matrix is generated based on the vector space model (VSM) to characterize the driving style fingerprint. Different from the traditional driving feature representation method, we propose a new concept of using bitmap to represent driving fingerprints. Given the sensor data captured by the vehicle during natural driving, we regard driving style recognition as a time series classification task and divides it into three classifications: normal driving, aggressive driving and drowsy driving. We evaluate our proposed model on the open natural driving behavior classification dataset UAH-DriveSet. Compared with traditional classification methods, our proposed model achieves the most advanced results on UAH-DriveSet, achieving 83% and 95% F1-measure scores in motorway and secondary road, which is more than 20% higher than the closest comparison method.

A detail-aware multi-angle vehicle recognition algorithm is proposed to address the problem of detail information loss due to pooling operation in multi-angle vehicle recognition. Firstly, considering that the differences between different vehicles are concentrated in the vehicle length and axle regions, the mid-level features are selected to build a local perception module, and the ECA attention mechanism is embedded to enhance the network's discrimination of details in local regions and suppress the interference of low discrimination features; secondly, a void space pyramid pooling model is introduced to increase the perceptual field to aggregate multi-scale contextual information, which reduces the problem of detail information loss and improves the global The second is the introduction of the null space pyramid pooling model, which increases the perceptual field to aggregate multi-scale contextual information, reduces the loss of detail information and improves the perception of global features. The experimental results show that the detection accuracy mAP of the improved algorithm reaches 96.1%, and the proposed method can effectively obtain local features with discrimination and improve the perception ability of the global features.

Emotion recognition in children with autism has been a hot topic of research at home and abroad, and research on emotion recognition through facial features has made corresponding progress, but the observation of facial features is somewhat subjective and less objective. The EEG signal was reconstructed by wavelet packet decomposition to remove noise and artifacts. After that, the frequency band energy of the four rhythmic waves was extracted as features, and the SVM model optimized by genetic algorithm and the unoptimized SVM model were subjected to control experiments. The average accuracy of the parameter-optimized model was 80.37%, which was higher than that of the control group at 64.12%, and the accuracy rate was improved to a certain extent, which can provide an effective and objective basis for the diagnosis of autistic children.

The aesthetic evaluation of handwritten Chinese characters has important due value in primary and secondary education. This paper proposes to use deep learning to solve this problem. Firstly, using the public handwritten Chinese character aesthetic scoring dataset, training is carried out in various deep learning network models, and finally the network model with better scoring effect is determined. The experimental results show that the deep learning network model used in this paper has an accuracy rate as high as 83% in the aesthetic evaluation of Chinese characters, and the error performance is better than that of human evaluation.

In order to reduce the large amount of storage space occupied by model pose library in part pose recognition, a storage optimization method based on region of interest is proposed. First, perspective projection is used to render part CAD model, and sampling on view sphere uniformly to get model pose library. The region of interest of pose image is then cropped to optimize pose library storage space. The region of interest is identified as follows: transform pose image into gray image and then black and white image, mark largest connected domain in black and white image and crop it with external rectangle. Then, the part real image is obtained and the region of interest is cropped similarly. The part real image is matched to the pose library by template match. Finally, the optimal match image is recognized by template match algorithm and applied to adjust part CAD model. Experimental results show that the average storage optimization rate of the two models used in experiment reached 88.2%, and the average time of parts pose recognition is about two seconds.

Convolution Neural Network (CNN) is an algorithm that takes an image as input, and then assigns weights and deviations to all aspects of the image to distinguish each other. The neural network can be trained by using a batch of images. Each image has a label to identify the true nature of the image. This paper gives the basic flow of CNN image recognition technology, and verifies the accuracy of its image recognition with the method of cat and dog image recognition. We used 1000 images for testing, and the recognition accuracy rate was as high as 94.5%, higher than the accuracy rate of 90.3% of SVM and 89.5% of KNN. The experimental results show that the image recognition technology based on CNN is a good image recognition technology, which can be widely used in real life.

Printed circuit boards (PCB) are manufactured and transported and stored in such a way that many factors can lead to different types of defects. Currently, manual defect detection and machine vision-based defect detection methods have problems such as slow detection speed, high false detection rate and fewer types of defects that can be detected. In this paper, a modeling method for PCB defect detection model based on deep learning is proposed. First, to address the problem of difficult feature extraction due to the small size of PCB defects, the original image is first generated at super-resolution by generative adversarial networks, and then the original backbone network of FasterRCNN is replaced by Restnet101 for feature extraction, with a deeper network ensuring better feature extraction results. Experiments show that the model can effectively detect six types of PCB defects: missing holes, pseudo-copper, short circuit, stray, mouse bite and open circuit.

For the requirement of automatic recognition of traffic police gestures in complex backgrounds based on vision sensors for driverless cars, we propose a method for traffic police gesture action recognition based on two-stream spatio-temporal attention graph convolutional network (2s-AGCN) with two different dimensional skeletal data. Firstly, detect the commanding traffic policeman in the video, extract the 2D and 3D skeletal data with the pose estimation algorithm to reduce the influence of complex background and joint overlap on action recognition, then, build the spatio-temporal graph model ; After that, we construct a 2s-AGCN network, input 2D and 3D skeletal sequences into the network to learn the spatio-temporal features of gesture actions. Finally, a fusion of the two-stream information is done and then output the final traffic police gesture category. 2s-AGCN uses Non-Local and TopK at the spatial level to focus on all nodes directly, selecting the strongest K neighbors of interaction strength; Temporal attention is used to focus on the frames that have higher contribution. The ablation study is done on the dataset CTPGD, and the results show that the method significantly improves the recognition accuracy of traffic police command gesture actions, especially those with overlapping skeleton points.

Facial Expression Recognition (FER) is a hard work, but in the current research on expression recognition technique, the regression of recognition performance is caused by misclassification problem which is caused by non-saliency of local features for facial image. In this paper, we introduce a framework for FER using inter-class separability for regional-feature fusion based on multi-domain expression generation (GSER).We fuse the regional features from multiple layers, and the local feature from most informative regions is enhanced to improve the precision of expression classification. An inter-class separability method based on pre-defined evenly-distributed class centroids is discussed for loss function, which solve the misclassification problem of expressions. The framework is evaluated over different standard facial expression datasets including KDEF and RAF-DB, and experiments show that the problems for unbalance of sample size and misclassification are solved, the robustness and effectiveness of expression classification is improved.

To address the time-consuming and laborious task of picking up soccers, an intelligent soccer-picking robot is designed and implemented, which consists of three main parts: mechanical, electronic, and software design. The robot employs a four-wheel drive and hydraulic shock absorber structure to ensure fast and stable operation. The main body structure of the robot is built with aluminum alloy and PLA materials to reduce weight effectively. The Kendryte K210 chip is the main control chip to reduce power consumption and cost. A monocular camera is used to construct a map, and the YOLOX object detection model is used to detect various objects on the soccer field. The artificial potential field algorithm is improved to achieve optimal path planning for the robot. The robot is tested 80 times on a real soccer field, with a successful pickup rate of 92.5%, meeting the actual needs of soccer-picking and filling the gap in the soccer-picking robot market.

Aiming at the problem of low efficiency and low precision of vehicle logo positioning, this paper proposes an improved vehicle logo positioning algorithm VLL-Net based on Yolov4. First, the Yolov4 network structure is improved, and the loss function of Yolov4 is modified, and the ResNet residual network structure is added between the convolutional layers before each YOLO output layer to solve the problem that CNN will follow the network during training. The problem of gradient disappearance occurs in the deepening of the algorithm. Focal Loss is used to solve the imbalance problem caused by too many negative samples during the algorithm training process, making the whole algorithm more suitable for the vehicle logo positioning task. The experimental results show that the improved yolov4 network model improves the detection accuracy.

Prolonged and poor sitting postures are major causes of many musculoskeletal disorders, such as herniated discs, cervical spondylosis, back fasciitis, and scoliosis, and hence proper postures are essential for maintaining good health. Towards a pervasive and low-cost healthcare system, we in this study design and implement a sitting posture monitoring system, named iGuard, with a pressure sensor array. Specifically, we use five sensors to sense the sitting posture information and naturally annotate the training sensor readings over a while. Afterward, we segment the streaming data and extract a variety of time-domain and frequency-domain features from the segments to train a sitting posture recognition model. Finally, comparative experiments are conducted to evaluate the performance of iGuard in distinguishing normal, left-leaning, right-leaning, forward-leaning, backward-leaning, left-leg crossed, and right-leg crossed sitting postures. Particularly, we consider three different evaluation schemes, i.e., subject-dependent setting, subject-independent setting, and cross-subject setting. Experimental results demonstrate the effectiveness of iGuard.

The detection and recognition of traffic signs acts an extremely crucial role in driverless technology and assisted driving systems. In view of the low recognition rate of traffic signs, this paper improves the neural network based on the existing VGG16 network architecture, including adding BN layer, adopting batch normalization algorithm and adding attention mechanism. The GTSRB data set is enhanced by scaling and random rotation. The data is preprocessed by image graying and image normalization. In this paper, the model recognition rate reached 99.34%, which is of great value to enhance the recognition accuracy of traffic signs and promote the development of intelligent transportation.

To alleviate the influence of angle change and dressing style change on gait recognition, a gait recognition network GaitRL based on local gait relationship learning is proposed. First, the pedestrian gait image is acquired by the camera equipment, the acquired image is preprocessed, and the pedestrian data is aligned with the central axis of the human body. Secondly, the image is segmented and local feature extraction is performed. Third, learn the relationship between local features A_m, calculate the product between Am and the image sequence, and use it as the data feature of the gait. The experimental results show that compared with the GaitSet network, the recognition accuracy of the network is improved by 0.7%, 2.4% and 3% under NM, BG and CL conditions, respectively, which can better mitigate the influence of angle change and dressing style change on gait recognition.

Brain-inspired intelligence is recognized as one of the important research directions of artificial intelligence. It is also an important supplement to brain science and brain medicine. At present, the research of Brain-inspired intelligence still focuses on traditional computer science. There are still many difficulties to overcome on the way of brain inspiration and imitation. In this paper, based on the Spiking Neural Network, the traditional Spiking Neural Network is optimized by brain inspired method to establish a Brain-inspired cooperative control system. This system will provide a new idea for the development of humanoid robot and exoskeleton technology.

The surface integrity, rot degree, mechanical damage, rupture, scar and other states of sand sugar orange fruit can be reflected from the surface defects of sand sugar orange. The surface defects of sand sugar orange can be used as an important characteristic parameter to judge the surface integrity, rot degree, mechanical damage, rupture, scar of sand sugar orange fruit. By segmenting the background of the collected feature image, the surface defects of the fruit peel were detected, the common surface defects of the orange were identified, and whether the orange was ruptured or scarred was judged, and the classification was realized.

Weakly-supervised object detection (WSOD) is the process of detecting instances with bounding boxes in an image using only image level labels. Multiple instance learning (MIL) based networks are one of the fundamental frameworks of WSOD methods. To convert object detection to MIL proposals classification, proposal methods such as selective search windows (SSW) interpret the image as a bag of proposals. While WSOD reduces the requirement for datasets, the lack of instance-level supervision makes it challenging to constrain the proposals, which poses a considerable challenge to both the training process and the final detection accuracy. In order for training to proceed smoothly, the model often has to ignore targets with complex spatial overlap and obscure features, and focus on learning regions with distinct features. This can lead to serious problems of object missing and wrong detections. To address the aforementioned issues, this paper proposed a co-teaching structure to improve the adaptability of the model to complex scenes. This approach involves analyzing the different performance of different categories of proposals in the training process, training two deep neural networks with mutual supervision simultaneously, and controlling the back-propagation loss function while ensuring the learning ability of the original MIL network. Experimental results have demonstrated that this structure can effectively enhance the accuracy of detection and can be seamlessly integrated into other models that perform the same MIL processing.

With the development of deep learning in diagnosing autism spectrum disorder (ASD), multi-site resting-state functional magnetic resonance imaging (rs-fMRI) images have made excellent progress. However, there is a heterogeneous problem between the multi-site data caused by inconsistent data distribution. To address this problem, we propose a combining autoencoder and category-based low-rank domain adaptation (AECLR) method for multi-site ASD identification. The main idea is to extract non-linear features and alignment the distribution of these features. In the first stage, the unsupervised autoencoder is used to obtain the non-linear representation. In the second stage, the common structure between all domains was mined by the category-based low-rank constraints, which transform all source domain data and the target domain data into the common latent space and then the source domain data could be linearly represented by the target domain data. As a result, the ablation experiments of the AECLR method achieve independent performance and the AECLR method also obtain a satisfactory classification when compared with the state-of-the-art method.

As the construction of oil and gas pipelines continues to advance, the safety of pipelines has garnered increasing attention. Pipeline welds represent a crucial factor in ensuring the safe transportation of pipelines over the long term. Therefore, to guarantee the quality and safety of pipeline welding, it is necessary to inspect the welds of oil and gas pipelines. Currently, manual detection is the most common method for inspection. However, as the number of pipeline detection points continues to increase, manual detection has revealed issues such as low detection efficiency and the susceptibility of detection results to subjective judgments. To address these shortcomings, this study introduces machine vision technology to the detection and recognition of oil and gas pipeline welds, which holds significant guidance implications for the detection and recognition of oil and gas pipeline welds.

Object grasping based on artificial neural network is a research hotspot in the field of industrial robot, and effective grasping is highly challenging in the robot control process. In the traditional robot object grasping process, it is necessary to obtain and store the three-bit model of the grabbed object in advance based on template control grasping trajectory, which leads to the grasping scene that is relatively fixed and single, lacking the flexibility of grasping. This study constructs a lightweight convolution neural network combined with artificial neural network technology based on the RGB-D camera, and points at for known objects and unknown objects two categories to study algorithm of the robot grasping position detection, which provides effective technical support to solve the problem of grasping position detection and improves the flexibility of grasping.

This paper proposes a new model, PGB_GPT, for Chinese fine-grained sentiment analysis, which combines Bidirectional Long Short-Term Memory (BiLSTM), Graph Convolutional Neural Network (GCN), and Generative Pre-Training Model (GPT). Additionally, a multi-core plant intelligent model is introduced to extract comprehensive symbolic meaning and improve the precision and accuracy of sentiment analysis. PGB_GPT outperforms other combination models and the possibility of merging a multi-core plant intelligence model with BiLSTM, GCN, and GPT for more extensive and accurate emotion analysis is highlighted. For Chinese fine-grained sentiment analysis, the PGB_GPT model combines BiLSTM, GCN, and GPT, with "P" representing "Plant Intelligence," "G" representing "Graph Convolutional Neural Network," "B" representing "Bidirectional Long Short-Term Memory," and "GPT" representing "Generative Pre-Training Model." As evidenced by the sentiment analysis dataset evaluation, each component greatly contributes to the model's enhanced performance.

With the advancement of artificial intelligence and image processing technology, hand-drawn sketches can be utilized in diverse applications, such as digital creativity, design, and education, among others, highlighting their vast potential in various fields. However, existing challenges in hand-sketch recognition such as low accuracy and efficiency necessitate the development of an improved recognition method. To address these challenges, a sketch recognition method based on Sketch-AlexNet is proposed. Specifically, a larger first layer convolutional kernel is selected to enhance feature extraction ability, while a smaller step size is used to minimize information loss. Moreover, a set of multiplicative convolutional kernels is introduced to replace the original convolutional structure, enabling the network to extract features across shallow to deep hierarchies. The Sketch-AlexNet model is trained on a self-built hand-drawn sketch dataset, yielding a 9.8% accuracy improvement, better recognition results, and corresponding advancements in recognition speed and stability.

With the progress of intelligent computers, the auto speech recognition function has developed rapidly. However, at present, the automatic recognition efficiency of on-board speech interaction is low and the recognition rate is poor. Even some vehicles do not display speech characters, which makes it impossible to verify whether the speech recognition is accurate. This paper proposes a multi-dimensional data method based on voice, picture data, text data, ADB data and internal changes of vehicle and machine, and improves the accuracy of vehicle borne voice recognition verification through multi-mode information fusion technology. Then, based on the parameters of speech wake-up rate recognition, false wake-up rate recognition, average wake-up time recognition, speech command response time recognition and speech command function recognition, a speech performance evaluation index model is established. Finally, the experimental results show that the verification method of multi-dimensional data speech recognition can verify the speech recognition rate and the execution rate of the vehicle in an all-round way.