This PDF file contains the front matter associated with SPIE Proceedings Volume 12782, including the Title Page, Copyright information, Table of Contents and Conference Committee lists.

Improved fuzzy c-means (FCM) clustering algorithms have been widely used for image recognition and localization. However, in industrial assembly systems, the unsatisfactory pixel merging and segmentation results between local adjacent windows, combined with the differences in the shape, size, and material of parts, as well as variations in lighting conditions, make target image recognition and localization a challenge. Most algorithms struggle to achieve the expected results and have high computational complexity. In this study, we propose a super-resolution-based FCM clustering algorithm that is faster and more accurate for image recognition and localization in industrial assembly systems with irregular part sizes. We first use multiscale morphological gradient operations to obtain high-resolution images. Then, we use the fast FCM clustering algorithm to achieve the recognition and extraction of specific target images. Finally, we use the Sobel operator to determine the target's position. The experimental results demonstrate that the proposed algorithm shows higher accuracy and efficiency in image recognition and localization for industrial assembly systems.

The Terracotta Warriors and Horses of the First Qin Emperor are a highly prized relic in China and the world. Some people believe that the Terracotta Warriors are a realistic representation of the Qin dynasty, while others feel that the Terracotta Warriors are the result of artistic re-creation. Therefore, in order to verify the "realism" of the terracotta warriors and horses, the degree of resemblance between the terracotta warriors and real people was quantified. The representative points of the outer corners of the eyes on the head and face of the terracotta warriors are selected. The method of extracting feature points based on the cross-sectional line of approximation is proposed using point cloud data. The method is based on the approximation method of cross-sectional line extraction, which is more accurate and easier to calculate by converting 3D into 2D. The final experimental results show that the curvature of the outer corner points of the terracotta warriors and the curvature pattern of the real eyes basically match. The method verifies that the facial features of terracotta warriors are highly correlated with those of real people. At the same time, the "realistic" nature of the terracotta warriors is demonstrated.

With the arrival of the Internet era, smartphones, tablets and other terminal devices have gradually become popular. On this basis, mobile applications based on the Android platform have emerged and developed rapidly, including two categories of multimedia UI design and embedded HTML5 webpage creation. In order to meet people's demand for information products, the two need to be effectively combined to achieve efficient information processing functions. At present, China has started to vigorously develop the intelligent operating system represented by Android system, and multimedia UI design is one of its important components. The work is mainly done by the software development team, so the developers are required to have a high level of professional skills. At the same time, it is necessary to ensure the flexibility and compatibility of the designed program. In addition, to make the computer technology better integrated into the multimedia UI design, it is necessary to prepare the corresponding data to ensure that the user can provide good visual effects and interactive experience effect.

This paper studies feature extraction and feature matching in visual odometry. Aiming at the problems that ORB feature extraction does not have illumination invariance and feature distribution is uneven, an adaptive threshold algorithm for feature extraction is added, and a quadtree is used to manage feature points. Aiming at the problem of high time cost of the feature matching algorithm, an outlier removal algorithm based on geometric constraints is proposed, and the constraint set is constructed by using the slope, distance, and descriptor distance between the matching feature point pairs. Tested on the TUM dataset, the feature extraction algorithm can adapt to scenes with different brightness, and the robustness is improved. The time taken by outlier removal algorithm based on geometric constraints is about 10% of RANSAC. After that, combined with RANSAC, the running time of RANSAC can be reduced by 60%. Our algorithm can improve the estimation accuracy and robustness of the system.

There are still challenges in the field of video understanding today, especially how to use natural language to describe the visual content in videos. Existing video encoder-decoder models struggle to extract deep semantic information and effectively understand the complex contextual semantics in a video sequence. Furthermore, different visual elements in the video contribute differently to the generation of video text descriptions. In this paper, we propose a video description method that fuses instance-aware temporal features. We extract local features of instances on the temporal sequence to enhance perception of temporal instances. We also employ spatial attention to perform weighted fusion of temporal features. Finally, we use bidirectional long short-term memory networks to encode the contextual semantic information of the video sequence, thereby helping to generate higher quality descriptive text. Experimental results on two public datasets demonstrate that our method achieves good performance on various evaluation metrics.

The current conventional train hook crack repair technology is mainly used to remanufacture and repair worn hooks by laser cladding repair technology, which leads to poor crack identification due to the lack of simulation and analysis of crack data. In this regard, a multimodal information fusion-based crack repair method for train hooks is proposed. The attention mechanism based on the attributes of multimodal information fusion is used to fuse the multi-scale image alignment method and calculate the crack image region features to realize the recognition of hook cracks. Based on this, numerical simulations of train hook crack repair are performed, and the repair process is optimized. In the experiments, the proposed method is verified for the crack recognition effect. The experimental results show that the proposed method has a high recognition accuracy and ideal crack recognition effect when the proposed method is used to recognize train hook images.

In response to the problem that the current image processing technology and underwater target recognition algorithms are not yet mature enough in the field of underwater archaeology, this article innovatively applies object detection and underwater image clarity technology to the field of underwater archaeology. We propose a method for detecting and recognizing underwater cultural heritage based on optical devices. The method includes ocean image preprocessing and underwater cultural heritage object recognition based on YOLO V4. The results of experiments demonstrate that the proposed method can effectively and accurately detect and recognize targets in the underwater cultural heritage scene, and the clear image of the underwater relics after image preprocessing can better assist archaeologists in observing the species and distribution of samples in the real scene.

The conventional discharge state identification method mainly focuses on partial identification. The field identification environment is subject to various interference signals from Getang, resulting in poor performance of the ring main unit discharge state identification. Therefore, an intelligent ring network cabinet discharge state recognition method based on audio signal spectrum analysis is designed. Collect the partial discharge data of the intelligent ring network cabinet, and extract the characteristics of the partial discharge of the intelligent ring network cabinet. Based on the audio signal spectrum analysis, the partial discharge noise signal of the ring main unit is processed, and the discharge noise signal is filtered to ensure accurate identification of the discharge signal. By means of comparative experiments, it is verified that the recognition effect of this method is better and can be applied to real life.

In order to plan and design different models of urban landscape, this paper presents numerical methods and calculation methods for different models of urban landscape. Through the analysis of urban green space landscape pattern, it can be found that there are still problems such as the incompatibility of green space and space distribution, the dispersal of some green space, the high fragmentation space, and the urban green space shortage. In the future urban green space development, attention should be paid to the following problems: there is less green space in the central part of the city, and the green space distribution is focused on the poor areas, especially in the northwest part of the city. In future urban development, especially in the context of old urban development, it is necessary to strengthen the construction of green space in the central region, appropriately provide some economic benefits, and increase the green space in the central region of the city by destruction, the construction of small parks and green space, so that the public can live together. Urban ecological green space is insufficient, and construction should be strengthened. City A has good natural resources, among which the ancient Yellow River and the Beijing-Hangzhou Grand Canal pass through the city, and the urban water network is developed. We should focus on building ecological green spaces on both sides of the river and reduce the construction projects of residential quarters within a certain buffer zone. At present, there are many other green spaces in urban green space, most of which are unused construction land or green space, which may be occupied in the process of urban development. How to ensure the quantity and quality of such green spaces will be the key issues for urban builders.

In the context of mobile Internet, time series analysis has become an important way to capture the characteristics of data such as periodicity and correlation. Establishing a temporal sequence analysis model as an effective means to capture data features, for the problems of irregularity, nonlinearity, and inconspicuous feature relationships that commonly occur in sequences. In this paper, we use convolutional neural network to extract the potential features in the sequence, and combine the long and short term memory network to analyze the temporal features in the data; meanwhile, due to the "gate" structure of the long and short term memory network, some noise in the data is introduced into the model for training, resulting in the overfitting problem. -The decode-reconstruction network structure is used to remove this noise and improve the accuracy of the model. In this paper, we use the stock data of CBS as an example and compare it with the existing algorithm model, based on which we demonstrate the higher accuracy of this algorithm with different domain data sets.

Design a three-dimensional digital intelligent patrol system for substation based on digital twin technology, and intelligently patrol the substation equipment to effectively excavate the potential safety hazards of the substation. The physical entity layer uses cameras and sensors to collect the image of substation equipment, parameter data and environmental information, and carries out effective preprocessing of the acquired data; The digital twin virtual model layer calls the relevant data of the physical entity layer, and uses SolidWorks, 3D MAX and Unity3D software to build the digital twin virtual model of the substation; The application layer plans the patrol path and identifies the electronic tag of the equipment according to the substation data sent by the network transmission layer. On this basis, the patrol module applies the Super SAB neural network equipment state perception and prediction method to effectively perceive and predict the status of the substation equipment, and visualizes the patrol results on the human-computer interaction layer. The experimental results show that the system can effectively inspect the substation equipment, present the inspection results through virtual vision, intelligently inspect the substation equipment, and apply it to practical work, which can achieve better intelligent inspection results in the substation.

In this paper, the non-cultural material heritage mobile terminal APP interface is the carrier, according to the user experience of the interactive interface design. By using user interview, observation, qualitative research and quantitative research, and based on the theoretical model of user experience, the author conducted data collection and analysis using user interview and questionnaire survey to obtain four evaluation indexes and eight sub-criteria for users' interaction interface of intangible cultural heritage apps. The analytic hierarchy process was introduced into weight calculation. The weight of each evaluation factor is obtained through investigation and calculation, and the evaluation level of each element is determined by referring to the Likert scale. The evaluation data of the design scheme is obtained through the questionnaire method, the fuzzy analysis is carried out on the results of the questionnaire, and the final evaluation results are obtained according to the principle of full membership to provide implementable improvement suggestions for the interactive interface design to improve the user experience. The research results have theoretical guiding significance for the interactive interface design of intangible cultural heritage apps.

To address the problems of insufficient accuracy and difficulty of application in the current Chinese image description field, this paper proposes an evaluation method based on semantic constraints in the target domain. Unlike previous research, this method acts on the output stage of the model, and based on the extraction of key semantics in the target application domain, it is constrained by the macroscopic semantic space of that domain or by introducing external semantic information from other visual tasks. The experiments show that the proposed method effectively improves the semantic coherence between the model output description sentences and the input images in the target domain, and is helpful for the practical application of image description in specific domains.

Image super-resolution reconstruction is an ill-posed problem, as a low-resolution image can correspond to multiple high-resolution images. The models SRCNN and SRDenseNet produce high-resolution images using the mean square error (MSE) loss function, which results in blurry images that are the average of multiple high-quality images. However, the GAN model is capable of reconstructing a more realistic distribution of high-quality images. In this paper, we propose modifications to the SRGAN model by utilizing L1 norm loss for the discriminator's loss function, resulting in a more stable model. We also use VGG16 features for perceptual loss instead of VGG19, which produces better results. The content loss is calculated by weighting both the VGG loss and MSE loss, achieving a better balance between PSNR and human perception.

Due to the complex and harsh oilfield production environment, oilfield enterprises need a set of digital intelligent monitoring system to realize the real-time monitoring and control of oilfield water injection. This paper designs a field monitoring system of oilfield water injection based on configuration software. The system uses the general monitoring configuration software of Beijing Force Control Yuantong Technology Co., Ltd., which can realize modular function division and has the functions of fault alarm data recording and query. The PID control technology is applied to the system, and the performance of the technology is verified by simulation. It has been verified that the system can control the flow stably and intelligently, and can ensure long-term effective operation.

Single-sensor multispectral imaging technology has been widely used in computer vision, mechanical diagnosis, cultural history protection and other industries due to its convenience and low cost. Single-sensor multispectral imaging can only generate a single mosaic image, so an efficient method is needed to convert mosaic images into multispectral images. Based on the concept of pseudo-panchromatic image, a 9-band multispectral imaging system is designed in this paper. We directly estimate the pseudo-panchromatic image from the mosaic image and use the correlation between the pseudo panchromatic image and each channel to generate a multispectral image by guided filtering and residual interpolation. The experimental results show that the multispectral images obtained by our method are superior to the other two methods in objective and subjective evaluation.

Image-based automatic identification of the cotton and flax fibers is extremely significant for the content quantitatively assaying in the textile industry. In this paper, a fiber identification method based on multi-scale features of the texture and Gaussian Process Classification (GPC) is proposed. Firstly, the images of the fibers are collected by an optical microscope and a set of image preprocessing approaches including image enhancement, local binarization, morphological processing is utilized to extract the fibers from the background. Next, the single fiber images are analyzed by the Discrete Wavelet Transform (DWT) and obtain the multiple-scale features of the texture. Then, the Gray Level Co-occurrence Matrix (GLCM) is applied to describe the spatial distribution features. Subsequently, extract the statistical feature from the GLCM and obtain a 42- dimensional feature vector that contains the fiber texture. Finally, 2610 images are randomly divided into train set and test set, and the recognition expert system based on the GPC is trained and validated accordingly. The test results on the test set showed that the classification precision - recall for cotton and flax fibers reached 96% - 97% and 97% - 95%, respectively. The method proposed in this paper can help workers quickly identify cotton fibers and flax fibers for further work, such as calculating the blending ratio of blended fabrics.

The end-to-end learning framework for autonomous cars has attracted considerable interest in recent years from both industry and academia. To control an autonomous car, this approach requires developing a model that links sensory perception, such as picture frames from a camcorder, to learn the optimal and safe driving actions. To develop a completely automated driving, it is possible to use professional driver experience more efficiently by using the end-to-end learning approach rather than spending a significant amount of money on labelling like bounding boxes for things. The conventional self-driving system, which typically consists of detection, positioning, mapping, and route planning manually fails to perform well with dynamic autonomous vehicles. We proposed a Convolutional Neural Network (CNN) framework for learning the autonomous driving to convert front-facing camera data from raw pixels and learn the steering commands. This end-to-end strategy has shown very promising results compared to the traditional self-driving models. The proposed CNN framework learns to intelligently drive in traffic on local roads with or without lane markings, on motorways, and with the least amount of training data from experts. Additionally, it also works smartly in places with obscure visual cues, including parking lots and gravel roadways. The system spontaneously picks up internal demonstrations of the essential processing phases, like identifying convenient road objects, by learning the human steering angle as the training input.

With the rapid development of civil aviation industry, the air traffic flow is increasing, which brings a large load to air traffic control, airports and other units, the safety of flight activities has become a research hotspot, flight conflict detection is a necessary link to ensure the safety of flight activities, the increase in air traffic flow requires its more accurate, efficient and stable operation. Based on the least squares support vector machine (LSSVM) in machine learning, this study uses the information provided by ADS-B, such as heading, position and altitude, combined with the regulations and conflict protection zones in actual operation, to classify the occurrence and severity of flight conflicts under the same moment, i.e., to perform multiple classifications, and uses a hybrid optimization algorithm of genetic + particle swarm to optimize this support vector machine model, and proposes A very efficient and accurate real-time flight conflict detection model is proposed. Finally, simulation analysis shows that the support vector machine is faster and more accurate than the traditional SVM, and has excellent conflict detection capability, and by differentiating the classified conflict levels and performing supervised learning, it can provide accurate warnings for upcoming flight conflicts, which can draw early attention of ATCs and provide a basis for the next flight conflict resolution. Eventually, the conflict detection model is expected to be compatible to airborne and ground surveillance equipment, which can significantly improve the safety of flight activities and has a broad application prospect and important research value.

In order to solve the multi-vehicle mutual occlusion problem encountered in 3D target detection by self-driving vehicles, this paper proposes a monocular 3D detection method that includes dynamic occlusion determination. The method adds a dynamic occlusion processing module to the CenterNet3D network framework to improve the accuracy of 3D target detection of occluded vehicles in the road. Specifically, the occlusion determination module of the method uses the 2D detection results extracted from target detection as the occlusion relationship determination condition, wherein the method of changing the occlusion determination threshold with the depth value is introduced. Then the occlusion compensation module is used to compensate and adjust the 3D detection results of the occurring occluded vehicles, and finally the 3D target detection results are output. The experimental results show that the method improves the accuracy of both vehicle center point detection and 3D dimensional detection results in the case of long-distance continuous vehicle occlusion. And compared with other existing methods, the accuracy of 3D detection results and bird's-eye view detection results are improved by 1%-2.64% in the case of intersection over union of 0.5. The method can compensate for the occluded vehicles in 3D target detection and improve the accuracy

This study introduces a number of enhancements to the YOLOX-S target detection network in an effort to address the issues of heavy traffic at the ferry, complex traffic environment, and sluggish detection speed. The conventional residual block in CSPDarknet, which has a significant number of parameters and high equipment requirements, is replaced by the MBConv module in the deep layer and by the Fuse-MBConv module in the shallow layer. This is completed for YOLOXS's backbone feature extraction network, CSPDarknet. The enhanced model's mAP value is 83.39%, 2.7% more than the baseline method. The experimental findings demonstrate that the enhanced method presented in this study is appropriate for the real-time detection of moving objects, such as cars and people, in the vicinity of the ferry entrance

Deep learning has achieved great success in many fields, such as image classification and target detection. Adding small disturbance which is hard to be detected by the human eyes to original images can make the neural network output error results with high confidence. An image after adding small disturbance is an adversarial example. The existence of adversarial examples brings a huge security problem to deep learning. In order to effectively defend against adversarial examples attacks, an adversarial example defense method based on image reconstruction is proposed by analyzing the existing adversarial examples attack methods and defense methods. Our data set is based on ImageNet 1k data set, and some filtering and expansion are carried out. Four attack modes, FGSM, BIM, DeepFool and C&W are selected to test the defense method. Based on the EDSR network, multi-scale feature fusion module and subspace attention module are added. By capturing the global correlation information of the image, the disturbance can be removed, while the image texture details can be better preserved, and the defense performance can be improved. The experimental results show that the proposed method has good defense effect.

Due to the characteristics of water and the particles in water, the underwater images will have problems of color deviation, low contrast and uneven brightness. Therefore, a statistical approach based on variance is proposed in this paper to correct color-biased images and then equalize the image brightness using the Gamma algorithm. Secondly, a local adaptive contrast enhancement algorithm improved by contrast code images and a restricted contrast adaptive histogram equalization algorithm are used to improve the contrast of underwater images. Finally, two different contrast enhanced images after color correction are fused by a multi-scale fusion algorithm to obtain high-quality underwater images. The results of the comparison experiments with some existing underwater image enhancement algorithms show that our algorithm can effectively improve the quality of underwater images.

Convolutional neural network-based target detection algorithms are widely used in vehicle detection due to their high speed and accuracy. However, existing algorithms are characterized by large computational volumes, complex network structures, and severe resource constraints. They make them difficult to be ported to mobile platforms and embedded devices. Therefore, the structure of the relevant target detection algorithm needs to be optimized to enable wider deployment of the algorithm. To address the problems mentioned earlier, a YOLOv5SCB lightweight target detection network model is proposed. In the presented model, Shufflenetv2 and CA module are introduced into the backbone network to reduce the complexity of the network model and improve the detection accuracy of the model. Furthermore, BiFPN is integrated into the neck network to improve the efficiency of network feature fusion and enhance the ability of network feature expression. The experimental data show that compared with the original YOLOv5, the model parameters of the proposed YOLOv5SCB are reduced by 62.4% and the overall detection accuracy is improved by 1.1%.

In order to show the transport conditions of goods on different roads and provide more real and three-dimensional transport information for situation inference users, this paper proposes a simple and PID controlled three-dimensional simulation method for truck formation based on Unreal Engine. Firstly, based on the basic theory of automatic control [1] , the longitudinal lollipop controller and the transverse PID controller are designed respectively based on the lollipop control and PID control ideas, and the perception-decision framework is combined to realize the automatic driving of the truck along the spline line on the road. On this basis, a truck controller is designed to realize the truck formation driving with high recovery degree based on the leader-follower strategy. The results show that the truck based on PID control can accurately drive along the road line. With the cooperation of truck formation controller, the whole process of formation, maintenance and driving of truck formation can be basically restored.

YOLOv5 is one of the target detection algorithms with fast detection speed and high accuracy, but it has the problems of insufficient sensory field and low accuracy of small target detection. In order to solve above problems, an improved YOLOv5 network model, i.e., an improved YOLOv5-TI model based on the attention mechanism, is proposed. The attention module is added to the backbone network when extracting features to improve the target detection accuracy, and the input features are shifted windowed for self-attention calculation to effectively utilize the features and improve the small target detection accuracy; the proposed model YOLOv5-TI is experimented on the self-built inland infrared dataset, and the mAP value reaches 95.5%, and the results show that YOLOv5-TI can effectively improve the target detection accuracy. The inland vessels equipped with visual intelligent perception system can effectively identify the targets on water, and they have wide applications in the fields of surface exploration and autonomous search and rescue.

With the advent of digital age, bill recognition has become a necessary work for many enterprises, institutions and individuals. This paper proposes a ticket recognition platform based on deep learning algorithm, which can automatically recognize various types of tickets and realize fast and accurate processing. The platform uses deep learning algorithms such as convolutional neural network, cyclic neural network and long and short time memory network and combines image processing technology and natural language processing technology to effectively solve the difficult problems in ticket recognition. The experimental results show that the platform has high performance in ticket recognition accuracy and speed and can meet the needs of practical applications.

UAV clusters are a new type of cluster mode, which should maintain electromagnetic silence as much as possible during the formation flight. When the deviation of the UAVs position occurs during the flight, a pure direction finding method can be used to adjust the queue position of the UAVs. In this paper, we model and analyze the problem of position deviation for UAVs based on pure direction finding by optimizing theory and grid convenience method. Firstly, we establish an unbiased positioning model of the transmitter under ideal conditions. Then, we modify the model by equivalent conversion of deviation and establish the final biased positioning model of the transmitter. Finally, we simulate the actual UAV positioning situation through MATLAB simulation and verify the feasibility of our model.

At present, object detection algorithm using artificial intelligence technology plays an increasingly important role in the field of computer vision, and plays an extremely important role in such practical application scenarios as automatic driving, urban monitoring, national defense, military and medical assistance. Different from visible light imaging, infrared imaging technology uses detectors to measure the infrared radiation difference between the object itself and the background, overcoming the difficulty of low light intensity and realizing infrared object detection in the low-light scene. In this paper, the traditional infrared object detection algorithm for low light background and infrared object detection algorithm based on deep learning are reviewed, and the current representative classical algorithms are compared, and the characteristics of the model combined with the actual application scenarios are analyzed. Finally, the difficulties and challenges that the current infrared object detection task facing are described, and the research direction of infrared object detection is prospected.

According to the World Health Organization, the current global death toll from road traffic accidents is as high as 1.3 million annually. The main cause of road traffic accidents is poor road conditions, and potholes on roads are the most serious type of road diseases. Therefore, timely detection and treatment of road potholes is very necessary. This paper proposes a method based on the use of YOLOv7 deep learning model to detect potholes on the road. At the same time, CBAM attention mechanism and optimization of loss function are added on the basis of this method. Combined with the idea of transfer learning, the improved YOLOv7 network is trained. The final test results are significantly improved compared with other road potholes detection models. F1 score is 78%, Precision value can reach 85.81%, and mAP value can reach 83.02%.

Apples are susceptible to diseases during growth that can reduce yields and cause economic losses. The common types of diseases of apple leaves are mainly spotted leaf drop, brown spot, grey spot, tobacco leaf blossom and rust. In this paper, a control system for the fine breeding of apple trees is designed to address the problem of the five diseases mentioned above affecting the growth of apple trees. The system uses a convolutional neural network to build a CNN model for disease identification of apple leaves. The data set is first processed using a pre-processing model (Xception) and the processed data is loaded into the built model. The experiments show that the accuracy of disease recognition using this model is high, and that fine farming of apple trees can be achieved through the control system.

Achieving automatic sorting of garbage through a mechanical arm depends on accurate recognition and localization of garbage. In this paper, we propose a garbage sorting method based on an adaptive deep neural network. The method addresses the limitations of YOLOv5 object detection algorithm, such as the fixed number of anchor boxes and the inability of the feature fusion network to adjust according to the target scale. Our proposed method introduces an object detection algorithm based on an adaptive deep neural network. We use the adaptive K-means clustering algorithm to automatically determine the initial clustering center and the number of clusters, extract features of multiple scales using the feature extraction backbone network, and automatically adjust the structure and feature fusion times of the adaptive feature fusion network based on the clustering results of adaptive K-means. We test the proposed algorithm and YOLOv5 object detection algorithm on a self-made garbage classification dataset. The experiments demonstrate that our proposed adaptive deep neural network reduces the model parameters of YOLOv5 by 27.03%, improves the detection speed by 18%, and enhances the detection accuracy by 0.7%. Finally, we transplant the adaptive deep neural network to the garbage sorting platform and use it for real-time garbage sorting.

With the rapid development of high-speed railway, it is necessary to ensure the safety of railway running, and the computer interlock system is the key equipment to ensure the safety of railway running in the station. It is a real-time system with high safety and reliability, which needs comprehensive and strict testing before it is put into use. In order to ensure that the interlocking system can strictly complete the function of each part, the computer interlocking test is very important. In recent years, with the rapid development of deep learning and image processing technology, in order to further improve the test efficiency of computer interlocking system, this paper studies the result decision module of automatic interlocking test. Target detection algorithm YOLOv5 is adopted to realize the location and recognition of signal, switch and section icon on the interlocking interface.

In construction, coal mining, tobacco manufacturing and other industries, wearing helmets is crucial safety measure for workers, and the monitoring of helmet wearing plays a significant role in maintaining production safety. However, manual monitoring demands substantial human, material and financial resources, and will suffer from low efficiency and are error prone. Therefore, we proposed a lightweight real-time deep learning-based detection framework called YOLO-H, for automatic helmet wearing detection. Our YOLO-H model was developed on the foundation of YOLOv5-n by introducing the state-of-the-art techniques such as re-parameterization, decoupled head, label assignment strategy and loss function. Our proposed YOLO-H performed more efficiently and effectively. On a private dataset, our proposed framework achieved 94.5% mAP@0.5 and 65.2% mAP@0.5:0.95 with 82 FPS (Frames Per Second), which surpassed YOLOv5 by a large margin. Compared to other methods, our framework also showed overwhelming performance in terms of speed and accuracy. More importantly, the developed framework can be applied to other object detection scenarios.

The situation of the future naval battlefield will become more and more complex, and it will become a trend to develop various military auxiliary decision-making systems based on artificial intelligence and big data technology. This paper sorts out the key technologies of the auxiliary decision-making system based on deep reinforcement learning. On this basis, it proposes the construction method of the naval aviation sea-striking agent model, and completes the construction of the training framework with the combat deduction system as the environment. Finally, it summarizes and prospects some of future work.

Aiming at the problem of low detection accuracy in traditional UAV target recognition, an improved YOLOv5 target recognition method is proposed. The loss function of YOLOv5 is improved, and the CIoU loss function is used instead of the GIoU loss function used by YOLOv5 to optimize the training model. The accuracy of the algorithm is improved, and a more accurate identification of the target is realized. The experimental results show that the mAP value of the model trained on the aviation dataset NWPU VHR-10 by the improved YOLOv5 algorithm reaches 93.33%, which is 4% higher than the original YOLOv5 algorithm.

Vision technology plays an important role when AUVs (Autonomous Underwater Vehicles) operate underwater. In this paper, the three-dimensional mode of binocular stereo vision is constructed to complete the positioning of the target. This article then explains the problem of distorted underwater images and introduces a method to correct distorted images. Based on underwater physical imaging, the underwater image processing methods are divided into underwater image enhancement and underwater image restoration. The research status of the two ways is analyzed and reviewed. The advantages and disadvantages of the above methods are summarized and discussed, and the future development trend is predicted.

The continuous advancement of artificial intelligence technology has made autonomous driving possible. However, duo the lack of sufficient data to train a good deep learning model, the current smart driving system can only rely on the driver for autonomous control, which may have serious consequences in the event of an accident. In practical applications, smart driving systems not only need autonomous driving technology, but must also be able to recognize obstacles and accurately avoid them without relying on manual manipulation, making the integration of autonomous driving features into vehicles a very promising research direction. To address this problem, we propose a novel segmentation method, AU-Net, which is capable of achieving accurate and complete segmentation of complex scenes by introducing an axial attention mechanism. We evaluate the performance of our model on the dataset Camvid, which improves 0.54%, 0.47%, 0.32% and 1.54% in the miaou, accuracy, percision and recall metrics, respectively, and the results show that our model is well adapted to complex scenes in intelligent driving detection.

With the underground mining in Yanghuopan mining area, the original stress balance state in the rock mass is broken, causing the rock strata and even the ground surface around the goaf to move and deform, resulting in land subsidence and deformation and damage of surface buildings, affecting industrial and agricultural construction and people's living environment. The application of small baseline Radar Interferometry Technology in land subsidence monitoring provides a new means for the monitoring and analysis of land subsidence in Yanghuopan mining area. In this manuscript, SBAS-INSAR technology is used to monitor the land subsidence caused by underground mining in Yanghuopan mining area. Based on 33 sentinel-1A images from June 2019 to August 2020, the surface deformation center, deformation rate, cumulative deformation variables and other information of Yanghuopan coal mine were obtained, and the surface deformation of the mining area was interpreted and analyzed. The main settlement area of the mining area is located in the east of the mining area. The maximum settlement rate in the mining area is -96mm/y, and the maximum cumulative deformation is -119mm.

To address the problem of slow inference in the original OpenPose pose estimation model and lower the computing power of the model, this paper first uses MobilenetV3 as backbone to make a lightweight improvement for OpenPose's network, followed by using label fusion correction to further improve the accuracy of the model. These steps make a real-time pose recognition system built on embedded devices on robots possible. The performance of the improved model is verified on the COCO dataset, and the results show that the accuracy of the improved model is not much different from the original OpenPose model, but the detection speed is improved by a factor of 4. Finally, a pose recognition model was trained on the self-built dataset using the skeleton map output from the improved model and validated on the test set, and the experiments indicated that the accuracy of the pose recognition model was 92.5%, which was real-time and suitable for various application scenarios.

With the rapid development of multimedia technology and the Internet, the security of multimedia data is facing more and more threats. Digital watermarking technology can effectively solve this problem, and compressed sensing theory is widely used in the field of digital watermarking. As the theoretical basis of the new generation of information processing, this theory can effectively solve the problem of multimedia data being attacked. Based on compressed sensing theory, combined with discrete wavelet transform and SVD decomposition, this research realized the embedding and extraction of watermark in the observation domain after image thinning, and tested different forms of watermark attacks such as salt pepper noise, Gaussian noise and filtering on the watermarked image. The experimental data shows that the PSNR value and NC of the algorithm are in line with expectations, and it has good robustness, transparency, and extraction resistance.

Images, being significant carriers of memories and information, are valued by people. To restore images, it is necessary to perform noise reduction processing to eliminate noise generated by camera equipment and other factors. Traditional denoising technology such as wavelet transform is used to help engineer restore a image. And in recent years, the introduction of convolutional neural networks has accelerated the progress of noise reduction research. Many classic models have been developed by researchers using U-shaped networks and other techniques. Researchers often use multi-scale approaches to obtain multiple feature maps and enhance their network with these features. Our work enhanced denoising network by introducing large convolutions, small convolutions, and Fast Fourier convolutions to capture feature information at different scales. Additionally, we used an SE block to introduce attention mechanisms into the network. As evidenced by experimental results, our network achieved outstanding performance.

Currently, lumbar spine diseases are becoming increasingly young, and with the aging of the population, clinical doctors are facing increasing pressure in detecting lumbar spine diseases. Therefore, an AI-based diagnosis system for lumbar spine diseases using nuclear magnetic images (MRI) has become a sustainable solution for early diagnosis. However, a large amount of work has shown the fragility of neural networks in unseen data distributions. Therefore, this paper proposes an adversarial training-based robust diagnosis method for lumbar disc herniation to address the fragility issue of deep models under specific small perturbations. By enhancing the robustness of the model to specific perturbations through adversarial training, the deep network can correctly classify lumbar spine MRI data with perturbations. The deep network model uses ResNet50, with adversarial examples containing adversarial perturbations added during training, followed by joint training of normal and adversarial examples, and Mixup augmentation from the perspective of data augmentation to further enhance the model's robustness. Through 5-fold cross-validation training, this method was verified to significantly improve the robustness of the model under adversarial perturbations (average recognition accuracy increased from 50.14% to 71.07%), while maintaining high recognition accuracy for normal samples (our method/baseline: 89.14%/89.05%).

The accuracy of tourists traffic prediction plays a critical role in scenic area management. Traditional methods of forecasting tourist attraction traffic rely heavily on static historical data, often ignoring important factors that affect the flow of tourists. This process is usually time-consuming. However, with the emergence of deep technology, it is now possible to use real-time data collection and analysis to design a temporal and spatial representation of data sources. And a deep learning-based tourist flow recognition model combined with dynamic time-bending distance indicators and a temporal feature recognition method with temporal data clustering analysis is designed. The method can use location big data to analyze traffic temporal types and identify traffic spatial distribution features, and the analysis results can help traffic and facility management in scenic areas.

Football is a high-intensity sport that requires high physical fitness from athletes. Physical fitness is a key element that every football player must master. This article aimed to design an information technology based football player physical fitness video monitoring system, which can provide coaches with detailed information about the athlete’s physical condition and help them develop training plans. This article mainly used experimental design and data comparison to analyze the pre and post monitoring data of football players. The experimental data showed that the speed standard deviation of athletes before and after monitoring was below 0.1 at 10m and 5m. A football player physical fitness video monitoring system based on information technology is a feasible way to provide coaches with detailed information about their physical condition and help them develop training plans.

Since the 21st century, artificial intelligence has been continuously evolving in various fields, particularly in agriculture. Vegetables, as a critical component of agriculture and human diets, have always been a focal point in terms of cultivation, production, and sales. Compared to traditional vegetable classification that requires professional knowledge and experience, AI technology utilizes computer vision to achieve automated classification. This study presents a deep learning-based vegetable recognition system aimed at automating the identification and classification of vegetables. The system utilizes a convolutional neural network (CNN) as its fundamental algorithm, integrating the traditional CNN architecture, which comprises convolutional layers, pooling layers, and fully connected layers. In comparison to other vegetable recognition systems on the market, this system utilizes a simpler architecture for processing and classifying vegetable images, significantly improving the accuracy and compatibility of identification. The research steps comprise data collection, image preprocessing, model training, and model testing. Experimental results demonstrate that the system can rapidly and accurately identify and classify various vegetables, with an average accuracy rate exceeding 95% on the test dataset, showcasing high practical value.

The GrabCut image segmentation algorithm based on the principle of graph theory has been extensively used in the field of computer vision. However, the shortcoming is that it requires human-computer interaction to complete the ROI region selection to solve the segmentation task of the foreground image. Therefore, it cannot meet the requirements of fully intelligent image processing. In order to eliminate human-computer interaction and realize smart region selection, this paper proposes a ROI smart region generating and fine-tuning method to improve the GrabCut method, so as to realize intelligent image segmentation. The experimental results show that our method is compatible with both single-target and multi-target foreground image segmentation solutions.

With the vigorous development of artificial intelligence and related technologies, domestic power generation enterprises have also promoted smart wind power projects. In this paper, a wind turbine condition monitoring system based on adaptive neuro fuzzy interference system (ANFIS) is proposed for the state perception of wind farms. A normal behavior model of ANFIS is established based on common monitoring and data acquisition (SCADA) data to detect abnormal behavior of captured signals and to indicate component failure or malfunction using predictive errors. At the same time, according to the theory of wind farm accident warning, this paper adopts the NJW spectral clustering method for the first time, and implements the group classification of wind field fans. Then, the Elman neural network model is adopted for any unit in a certain group, so as to determine the working conditions of all units in a certain group. This method can effectively improve the efficiency of wind farm accident early warning, and is of great significance for the development of intelligent wind field.

The advancement of remote sensing technology has broadened the application scope of remote sensing image data across various fields. Traditional methods, when processing remote sensing images, face limitations in efficiency and generalization capabilities due to their intricate geographical features. In contrast, deep learning segmentation methods exhibit superior performance but struggle with contextual detail loss and multi-scale features. In this paper, we introduce the RSFNet model to tackle these issues. The model employs spatial paths to extract detailed information from low-level features, presents a residual ASPP incorporating an attention mechanism, and utilizes a feature map slicing module to capture small target features. Experimental results show that RSFNet attains 88.38% pixel accuracy (PA) and 81.06% mean intersection over union (mIoU) on the Potsdam dataset, proving its suitability for semantic segmentation of remote sensing images.

Against the poor noise immunity of the vibration measurement algorithm based on video phase, this paper proposes an algorithm based on enhanced fast empirical mode decomposition and Hilbert phase-based motion estimation (EFEMD-HPME). EFEMD decomposes the multicomponent video image into a single-component image, and then the local phase information of single component image is extracted through Hilbert transform, which has superior noise immunity compared with the spectral decomposition technique of the band-pass filter. Experiments show that the algorithm proposed in this article has a signal-to-noise ratio improvement of about 30% and a relative error of less than 0.5% compared with the HPME, which is of great significance for improving the robustness of video vibration measurement in general measurement environments.

At present, the image processing of remote sensing technology mainly depends on the transcendental ability of human beings, and it needs to spend a lot of artificial resources to mark. Therefore, this paper proposes a research and application of semantic segmentation method for remote sensing images based on convolutional neural network. Normalize the data, subtract the mean value and divide it by the standard deviation to standardize, divide the data, introduce data enhancement to further enhance the training data, and create a convolutional neural network and a training network. Each layer of U-NET is composed of three layers of convolution, and features are extracted and integrated by pooling or up-sampling. At the last layer, all the previously extracted features are classified into two categories to realize the semantic segmentation of the image. The experimental results show that the F1 score, Recall score and Precision score of this method are 84.31%, 89.59% and 79.62%, respectively. By introducing U-NET, the semantic segmentation accuracy of remote sensing images is improved. Compared with the traditional full convolution neural network, U-NET has been improved. Through the stronger connection between layers, plus up-sampling and down-convolution, features can be fully extracted and accurate segmentation can be achieved with fewer training samples.

Accurate semantic segmentation of retinal images is very important for intelligent diagnosis of eye diseases. However, the large number of tiny blood vessels and the uneven distribution of blood vessels in the retina pose many challenges to the segmentation algorithm. In this paper, we propose a Hybrid Attention Fusion U-Net model (HAU-Net) for segmentation of retinal blood vessel images. Specifically, we use the U-NET network as the backbone network, and bridge attention is introduced into the network to improve the efficiency of vessel feature extraction. In addition, we introduce channel attention and spatial attention modules at the bottom of the network, to obtain coarse-to-fine feature representation of retinal vessel images, so as to improve the accuracy of vascular image segmentation. In order to verify the model's performance, we conducted extensive experiments on DRIVE and CHASE_DB1 datasets, and the accuracy reach 97.03% and 97.72%, respectively, which are better than CAR-UNet and MC-UNet.

To address the issue of high cost and low efficiency in the manual sorting of apples, we proposed an improved apple classification method based on the AlexNet architecture. The algorithm added a batch normalization layer after each convolutional layer in the network structure to speed up the model's training process. Furthermore, we replaced the fully connected layer with a global average pooling layer to reduce the number of training parameters and save model training time. To improve the algorithm's robustness, we also performed data augmentation on the training samples before validating the algorithm to obtain an expanded dataset. Experimental results showed that the improved AlexNet network shortened the training time by 0.54%, increased the testing speed by 2.5%, and improved the accuracy by 1.12% compared to the original AlexNet network. Moreover, the training time of the improved AlexNet network was lower than that of other networks (AlexNet, ResNet50, Vgg16). The improved AlexNet network can efficiently and quickly classify apples and promote the automation of apple classification.

The actual underground situation is of great significance for workover operation. Videolog visualization technology can clearly and accurately obtain the underground color video information, and provide effective guidance for workover operation. This paper introduces the system composition, working principle and functional parameters of Videolog equipment, and gives an example of its practical application in workover operation, which shows that Videolog visualization technology is more efficient, safe and intuitive than traditional downhole video technology, and has a good application prospect in workover operation field.