This PDF file contains the front matter associated with SPIE Proceedings Volume 11069, including the Title Page, Copyright information, Table of Contents, Author and Conference Committee lists.

Fluid phenomena such as smoke, fire, and waterfall can be easily acquired via a video recorder in the nature. However, only very few researches are about how to edit and reuse the fluid sequences. In this paper, we develop a novel method to edit the fluid sequence spatially and temporally based on grid deformation. We use an as-rigid-as-possible (ARAP) deformation model to edit the fluid sequences. ARAP is like energy model to deform a coarse regular grid in real-time editing stage. Our ARAP model penalizes three parts of energy to make the deformation as rigid as possible. This model is very easy to solve, and users can adjust the weighting coefficients for the three energies to obtain various results. In the rasterization stage, one can use linear interpolation to export the dense texture coordinate field for the grid. And finally we integrate our deformation method with some other techniques to achieve different types of editing effects.

We present a predictive-corrective semi-implicit incompressible smoothed particle hydrodynamics (SPH) approach. At the first stage of the explicit computation, the fluid density is predicted using a continuous density method, which takes the velocity as the source term and considers the relative velocity effect. At the second stage of implicit computation, the ratio of the pressure to the square of the density obtained by the predictive-corrective method is directly used to correct the fluid velocity. As there are larger density errors in fluid particles near or on the boundary, more iteration numbers are required. The boundary handling method is applied to modify the density of fluid particles near or on the boundary. For density invariant conditions, the relaxation method and optimal initial value are adopted to improve the convergence speed. Our proposed approach can achieve semi-implicit incompressible SPH, solve the particle deficiency problem and realize the simulation of small-scale ocean scenes. While ensuring stability, the approach allows large time steps, controls the density deviation below 0.01% and improves the overall efficiency.

Nowadays, the models of deep learning are increasingly used in various industrial applications. However, the storage space of the models is too large, which makes it quite difficult to apply to mobile devices. In order to solve this problem, a model compression algorithm based on optimal clustering is proposed in this paper. Firstly, the model parameters of fully connected layer in deep convolutional neural network are clustered according to the best clustering method. Then the cluster center of the parameters is selected as the representative of the original parameter matrix. At last, the parameters of the cluster center are transformed differently in the forward calculation of the model to achieve the effect of compressing the parameters of the model and ensured the accuracy of the model. The compression algorithm proposed here is compared with other model compression algorithms in several common deep learning models such as Alexnet, VGG16 and so on. The results show that the algorithm proposed in this paper can compress the memory of the model greatly and improve the accuracy.

In fluid simulation, the analysis and calculation of various forces within the fluid is critical, since the detailed motion of the fluid animated surface depends on the motion of the particles caused by various forces between the internal particles of the fluid. The vorticity constraint of the fluid is ubiquitous in the fluid, especially in the simulation of multiphase fluids, due to the different properties of the viscosity, density and other properties of the different phase fluids, resulting in more complex vortex generation. We mainly modify the vorticity method and the single-phase fluid simulation method in the previous grid simulation, optimize it, apply it to the simulation of multi-phase flow in the hybrid model to simulate the vorticity phenomenon, so that the fluid animations are more vivid and more realistic.

Correct tracks obtained through effective data preprocessing are important to multi-target data fusion, especially in the circumstance of dense targets and strong interference. Outliers and target switch arise in the uncertainty of measurement data in the practical applications. In this paper, we propose a novel real-time data preprocessing method for outlier detection and target switch identification to obtain correct tracks, named as IDCDP (i.e., innovation and density combined data preprocessing). Experimental results demonstrate the effectiveness of IDCDP, which achieves outlier-free adaptive filtering and provides authentic tracks for multi-target data fusion.

Support vector machines (SVMs) have been widely used for binary classification. But large-scale training set will bring huge computation to the SVM. Researcher have proposed many techniques to improve the training efficiency of SVMs, and a typical class of improved SVMs is based on sparsely reducing training samples. To achieve this, clustering-based methods are most commonly used. However, clustering-based methods are ready to be disturbed by noise points. In order to solve this problem, this paper proposes a robust and efficient SVM algorithm based on K-Medians clustering (REK-SVM). Here, for each cluster, the cluster center takes the median value of each dimension attribute in the cluster, which can reduce the noise points. Especially, when the number of noise points distributed discretely is less than half of the total number of samples in the cluster, noise interference can be completely removed. The noise-free or noise-reduced subset data is used to train the SVM model. Experimental results show that our algorithm is fast and effective. For the processing of noise-containing classification data, its performance far exceeds SVM in terms of classification accuracy and efficiency. Compared to the K-SVM, they have the same computational complexity, but our algorithm is much higher than K-SVM in classification accuracy.

Image retrieval plays a key role in finding the required online multimedia learning materials quickly. In this paper, we realize a high-performance image retrieval by using bitrate allocation. First, the global descriptor is extracted to represent each image of the database. Second, the intra-group distance and the inter-group distance for each image are computed based on the global descriptor. Then, the retrieval difficulty degree (RDD) is evaluated based on the two types of distances. After that, an accurate bitrate allocation scheme is proposed when building the feature database according to RDD: using a higher bitrate for images that are hard to retrieve and lower bitrate for simpler ones. Finally, the higher retrieval performance is achieved based on the constructed highly efficient database. Experiments demonstrate that our method far outperforms the state-of-the-art on image retrieval benchmarks.

Image Stitching with large parallax has always been a challenging task, and accurate image alignment is critical for stitching results. In this paper, an image stitching method based on superpixel segmentation regions is proposed. To solve the problem of insufficient matching feature points under large parallax, an improved multi-plane RANSAC method is used to improve the robustness of matching feature selection algorithm. In terms of image alignment, a mesh optimization method with the global similarity prior is adopted, and a superpixel-based segmentation method is used to obtain reasonable matching points and global similarity transformation parameters. A standard seam-cutting algorithm is finally used to compose images together. Experiments show that the proposed method can effectively improve the performance of image stitching in complex scenes with large parallax.

The task of visual odometry (VO) is to estimation camera motion and image depth, which is the main part of 3D reconstruction and the front-end of simultaneous localization and mapping (SLAM). However, the accuracy of most of the existing methods is low or some advanced sensors are required. In order to predict camera pose and image depth at the same time with high accuracy from image sequences captured by monocular camera, we train a novel framework, named MD-Net, and it is based on convolutional neural networks (CNNs). There are two main modules: one is camera motion estimator which is able to estimate the 6-DoF camera pose, the other is depth estimator computing the depth of its view. The keys of our proposed framework are that we can not only train our two independent estimators, but also predict depth and camera motion simultaneously. What’s more, our motion estimator includes some shared convolutional layers and is divided into two branches to estimate camera orientation and translation, respectively. Experiments on KITTI dataset and TUM dataset show that our proposed method can extract meaningful depth estimation and successfully estimate frame-to-frame camera rotations and translations in large scenes even texture-less. It outperforms previous methods in terms of accuracy and robustness.

This paper introduces a new method for background reconstruction. Background reconstruction from video sequences captured by a static camera can be regarded as a low rank factorization problem. Background is the low dimensional subspace restored from the higher dimensional visual data, and foreground is treated as sparse noise of unknown distribution. The existing algorithm could not deal with noise of unknown distribution effectively. Due to the limitation of the matrix decomposition which would lost space structure information, we process video data directly as higher order tensor based on low rank tensor factorization (LRTF). We put forward a new model of foreground by using Mixture of Gaussians (MoG) and Markov Random Field (MRF). Extensive experiments show that our method can effectively construct the background.

The unstable quality of 3D printing products is one of the obstacles that hinder the popular application of 3D printing technology. In this paper, we propose a novel mask projection 3D printing scheme with visual surveillance. We introduce a vision system to monitor the gray level variation, so that two kinds of problems could be detected in time: 1. the printed objects are stuck on the bottom of the tank. 2. the printed objects are stretched out and part of objects are stuck on the bottom of the tank. The normal and abnormal printing procedures correspond to different gray level variation curves, which can be discriminated by classify algorithms. By the computing capability of our computer, we could monitor gray level variation curves at 470 points in real time. If 95 % curves are normal, the layer is thought successfully printed, the next layer will be printed continuously. Otherwise, this layer is abnormal, and the printing will be stopped. The experiment show that the proposed scheme can stop the abnormal printing in time. it can improve the production rate of the finished products and reduce the material waste.

The manufacturing industry has gradually developed from mass production to personalized customization, which requires a highly flexible robotic automated production line in manufacturing to deal with rapid retrofits of products. Industrial robots need to be taught the steps of completing tasks when dealing with new tasks. At present, the traditional robot teaching reproduction has many disadvantages, such as the complicated teaching system, the high cost, the waste of time, and the special training that operators need. In view of the above situation, a kind of teaching-playback based on machine vision and human hand demonstration was proposed in this paper. The position and orientation of the human fingertip during the execution of the mission were obtained by the Kinect sensor, and the pose information was transformed into the robot base coordinate system by coordinate transformation. The coordinate information was then transmit to the robot to enable the robot to quickly grab and place the target object. The feasibility and effectiveness of the teaching-playback system were proved by the experiment of picking and placing cylindrical workpiece.

A novel video stitching algorithm is proposed which stitches video streams captured by multi-UAVs into panorama video streams. Each UAV captures one video streams, there are independent shakiness and parallax between each temporally aligned frames. A pure 2D homography does not have the capability to solve parallax like spatial-variant homographies, where each input video frame is sliced into grids and each grid is transformed with a local 2D homography. Video stitching has to consider video stabilization due to strong temporal correlations between frames. Video stitching and stabilization are combined, since they are both about putting the grids from video sources into panorama properly. Good stitching results are generated by optimizing the camera paths which describe the grid transformation between girds of frames before and after. In 4 aspects they are optimized: 1) spatially with the grids around, 2) temporally with the grid at the same place from before and after frames, 3) geometrically with the grids from another video, 4) reliably not too far from the original transformations. We focus on video stream processing which does not need the whole video frames, only 7 frames later than the input, state-of-the-art results are generated.

With the fast development and application of virtual reality (VR), panoramic images have been a hot research topic and been widely applied in many areas. Usually, the panoramic images are generated by image stitching with images captured by cameras set in 360 degrees. However, the image captured by the traditional fish-eye lens has serious distortion. Many traditional image fusion algorithms are not suitable for this kind of images. In order to handle this problem, we propose a method based on multi-band image blending to realize the real-time stitching of six-way fish-eye video, and finally generates output video. The principle of the stitching system is to decompose the video into frames. After image alignment, multi-band fusion method is used to fuse the high-frequency features, which makes the degree of fusion in the image area smaller and avoids duplication in the mosaic area. Different from the convolution pyramid stitching algorithm which does not depend too much on the overlap area, the proposed algorithm retains more image details and reduces the distortion. Experimental results in terms of both subjective and objective quality show that the system can stitch panoramic video efficiently, and the output image has no stitching artifact in the overlapping area.

Cascaded face detectors using convolutional neural network (CNN) are widely adopted in real-world applications due to their superior efficiency. However, the performance of those detectors degrades significantly with the presence of challenging faces. In this paper, inspired by the fact that large receptive field provides more contextual information for improved performance, based on the existing cascaded CNN face detector, we propose a context-aware cascaded face detector, which aggregates more information for each candidate window. We empirically evaluated different methods for enhancing the information embedded in the receptive field and provide detailed comparisons. Moreover, we propose a score fusion strategy to combine the confidence scores from different detecting stages in the cascaded network to get a more reliable confidence score. We also propose an improved strategy to perform non-maximum suppression that achieves better efficiency when the number of proposals is huge. Extensive experiments on FDDB and WIDER FACE datasets demonstrate that our real-time detector achieves better performance over other cascaded detectors.

Distance metric learning between pairs of samples is the key issue for the person re-identification(re-ID). Recently, convolutional neural networks (CNNs) significantly is employed to improve performance by learning deep semantic features which requires a large amount of labeled data for train the deep model. For person re-ID task, the training data is scarce. To the end, how to use the limited data to achieve the optimal deep model is a challenging problem. We observe that sample's similarity information and identity information are complementary. Therefore, a full-batch loss function is proposed in this paper, which promotes the distance relationship of individual sample pairs in a batch to the distance matrix in all samples to make full use of the sample similarity information. At the same time, it also makes full use of the sample's identity information by integrating the identification loss function. We conduct experiments on the two public person re-ID benchmarks: Market1501 and DukeMTMC-reID. The results clearly demonstrate that our proposed full batch loss model produces more discriminative descriptors for person re-ID, which outperforms well established baselines significantly and offer new state-of-the-art performances

Person re-identification is indeed a challenging visual recognition task due to the critical issues of human pose variation, human body occlusion, camera view variation, etc. To address this, most of the state-of-the-art approaches are proposed based on deep convolutional neural network (CNN), being leveraged by its strong feature learning power and classification boundary fitting capacity. Although the vital role towards person re-identification, how to build effective CNN baseline model has not been well studied yet. To answer this open question, we propose 3 good practices in this paper from the perspectives of adjusting CNN architecture and training procedure. In particular, they are adding batch normalization after the global pooling layer, executing identity categorization directly using only one fully-connected layer, and using Adam as optimizer. The extensive experiments on 3 widely-used benchmark datasets demonstrate that, our propositions essentially facilitate the CNN baseline model to achieve the state-of-the-art performance without any other high-level domain knowledge or low-level technical trick.

Outdoor target tracking UAV (Unmanned Aerial Vehicle), which is a research hotspot in the field of computer vision and unmanned aerial system, needs robust target-tracking algorithms with good real-time performance, accurate position estimator of UAV and the corresponding control strategy of the system. In this paper, we designed an outdoor drone tracking system using PCA (Principal Component Analysis) face recognition algorithm and KCF (Kernel Correlation Filter) target tracking algorithm. Firstly, an image acquisition unit is constructed by using an on-board pan-and-tilt camera to capture an outdoor monitored area. Secondly, the PCA algorithm is used for face matching, then the tracking mode is automatically transferred when the expected face target is recognized. Finally, the target tracking is performed by the KCF algorithm. After that, the position error is calculated and sent to the flight control system through the MavLink protocol, thereby performing posture adjustment and completing the tracking and monitoring task. Experimental results show that the performance of outdoor target tracking flight robot is stable and reliable, which meets the requirements of outdoor target tracking and has a good application prospect.

More training samples can represent more various changes of face image, owing to the variational illumination, facial expressions and poses. However, in real-world applications, due to the limitation of storage capacity and captured image time, the number of face training samples obtained are often limited. In this paper, for the small size problem, we propose to firstly generate various mirror images by the original face images, viewed as its mirror image. Its mirror image is opposite to the original face image with facial details. Then the original sample and its mirror image is respectively used to perform collaborative representation classification method (CSC). Finally, an adaptive weight selection method is proposed to fuse the original face image and its mirror image based on assigning a better weight to the original face image. The results of experiments show that the presented scheme is effective.

ID face photos are widely used for identity verification in many real-world applications. In these cases, the face from ID photo is expected to compare with the face obtained from the daily life image. To improve the ID/life face verification performance, the datasets containing various ID face photos and daily life face images are always needed to train an available CNN model for feature extraction. However, it’s really hard to get uniform distributed data for both the ID photos and daily life images. In most face datasets, each subject only contains one ID photo (for example, obtained from the identity card) while contains varies of life images. This imbalanced distribution causes the difficulty of learning a general deep feature representation for both the ID and life faces, since the daily life images tend to have a higher impact on the feature learning than the ID photos. To address this challenge, we propose a weighted center loss which aims to learn a center of deep features for each class and penalizes the distance between the image features and their corresponding class centers with different weights. By emphasizing the ID photo features in the loss computation, our weighted center loss can train a general CNN which obviously narrows the distance between the ID and life faces. Final experiments demonstrate that with the joint supervision of softmax and weighted center loss, our proposed framework can minimize the intra-class variations caused by the imbalanced ID/life data distribution and significantly improve the ID face based verification accuracy.

Sparse discriminative multi manifold embedding (SDMME) algorithm was used for feature extraction, graph construction and projection learning were independent, the quality of the graph directly affects the effect of projection learning. In order to solve the problem, a new algorithm named sparse discriminative multi manifold embedding based on graph optimization (GOSDMME) was proposed in this paper. First, in proposed approach, the image matrix was divided into blocks. The matrix blocks on the same image were located on the same manifold. Then, the sparse graph was used to establish the connection relationship between different blocks. Finally, in the framework of the same objective function, the sparse constraint graphs and projections were studied simultaneously. The graphs and projections were learned at the same time, iterate and update the graph and projection to obtain a projection matrix that satisfies the accuracy requirements. The face recognition experiments conducted on Extended Yale B and CMU PIE datasets show that the new algorithm has better recognition performance than the SDMME algorithm.

Person re-identification (ReID) is an important task in video surveillance and can be applied in various practical applications. The traditional methods and deep learning model cannot satisfy the real-world challenges of environmental complexity and scene dynamics, especially under fixed scene. What’s more, most of the existing datasets are outdoor and has a single style, which is not good for indoor person re-identification. Focusing on these problems, the paper improves a Stride Convolutional Neural Network (S-CNN) to process indoor images based on multi-features fusion. The deep model is established in which the identity information, stride information and other information are learned to handle more challenging indoor images. Then a metric learning method (Joint Bayesian) is employed based on the deep model. Finally, the entire classifier is retrained with supervised learning. The experiment is tested on the OUC365 dataset created by us which is captured for 365 days including all seasons style. Compared with other state-of-the-art methods, the performance of the proposed method yields best results

Person re-identification (Re-ID) is an object recognition method based on visual appearance information. It is mainly restricted by the changes in person posture, shooting angles, the changes in the front, back and light of people that are mainly captured and the noises caused by shake or blur. Currently, single-frame person Re-ID is still the mainstream research. In view of the limited information of single-frame images, this paper adopts temporal attention sequence modeling to conduct research on person Re-ID based on video sequences, considering not only the content information of images but also the movement information between frames, etc.

In this paper, a temporal attention quality aware network (TA-QAN) is proposed. By extracting the temporal information between frames, all the frame sequences in the complementary information are effectively aggregated, and the influence of the quality image region is significantly reduced. The temporal attention quality aware network is used to extract temporal information between frames through temporal convolution. The comparison experiment with other feature extraction methods shows that the proposed method has the best performance in PRID 2011 and iLIDS-VID2014 data sets.

Automatic classification of digital patent images is significant for improving the efficiency of patent examination and management. In this paper, we propose a new patent image classification method based on an enhanced deep feature representation. Convolutional neural networks (CNN) is novelly applied to the patent image classification. The synergy between deep learning and traditional handcraft feature is explored. Specifically, the deep feature is first learned from massive patent image samples by AlexNet. Then such deep learning feature is further enhanced by fusing with two kinds of typical handcraft features including local binary pattern (LBP) and adaptive hierarchical density histogram (AHDH). In order to obtain a more compact feature representation, dimension of the fused feature is subsequently reduced by PCA. Finally, the patent image classification is conducted by a series of SVM classifier. Statistical test results on a large-scale image set show that the state-of-the-art performance is achieved by our proposed patent image classification method.

This paper aims to highlight vision related tasks centered on “shrimps”. With the further study of computer vision of marine life, we show that “shrimps” has been largely neglected in comparison to other objects. In image classification, the degree of visual separation between different shrimp categories is highly uneven, the appearance of some categories in same genus is very similar, and it is more difficult to distinguish than others. Based on the classification model of traditional convolutional neural network, this paper presents a method of merging shallow and deep features extracting feature maps from different levels according to the characteristics of shrimp. In order to facilitate future shrimps-related research, we present our on-going effort in collecting a dataset in this paper, “ShrimpX”, that covers not only shrimps and lobsters living in the sea, but also some freshwater shrimps. The “ShrimpX” dataset contains a variety of shrimp images crawled from image search engines. Experimental results on the “ShrimpX” dataset demonstrate that the proposed method can effectively improve the accuracy.

Recently, discriminative object trackers based on deep learning have demonstrated excellent performance. However, the tracking accuracy is facing a challenge due to contaminated training samples and different complex scenarios. For this reason, we propose a tracker based on sparse robust samples and convolutional residual learning with multi-feature fusion (SR_MFCRL). First, a sparse robust sample set (SRSS) is introduced to improve robustness of the network. In this process, we first employ sparse representation to estimate the best candidate and then utilize joint detection with response peak value and occlusion detection to determine the contamination degree of the sample. Second, a multifeature fusion residual network (MRN) is proposed and its two base branches to capture response output of different features in order to achieve higher positioning accuracy. Extensive experimental results conducted on OTB-2013 illustrate that the proposed tracker achieves outstanding performance in terms of tracking accuracy and robustness.

ORB (Oriented FAST and Rotated BRIEF) algorithm is widely used in feature point matching with images. However, the randomness of the threshold of search strategy makes the matching result inaccurate. The matching result of ORB algorithm is lack of robustness. In this paper, we proposed an improved ORB algorithm based on PSO (Particle Swarm Optimization) algorithm. Firstly, ORB algorithm was used to detect image feature points. Secondly, distance similarity measurement is applied to ORB and orientation constraint was added to reduce mismatching rate. Finally, particle swarm optimization algorithm was used to optimize the threshold of search strategy. Experimental results showed that the improved algorithm can effectively improve the accuracy of image matching and expand the scope of application of the algorithm.

Aiming at the problem that the matching accuracy for the low-texture areas in the stereo matching is so low that the matching failure is likely to occur. This paper proposes a stereo matching algorithm for low-texture areas. Firstly, using the horizontal and vertical gradient of the pixel and the specificity of the pixel itself to perform the calculation of the matching cost. At the same time, introducing the HSV color space to obtain the adaptive support arm length of each pixel in the image, thereby obtaining the aggregate area of each pixel. Then, in the acquired aggregation area, by using the bilateral filtering method to calculate the aggregation cost. Finally, in the disparity computation and refinement stage, Left-Right Consistency (LRC) check is combined with the iterative guided filtering method to reduce the error matching. At the end of the article, by using the standard images on the Middlebury platform to experiment. The experimental results are compared with the traditional experimental results, which proves the effectiveness of the proposed algorithm in the low-texture areas.

Aiming at the problem that the traditional one - dimensional linear processing method can’t effectively extract the characteristic information of the weak signal of the brain, this paper presents a lesion location algorithm of electroencephalogram (EEG) based on image features. Image features is introduced into epileptic signal processing, which can extract feature information and locate lesions in EEG weak signals. First of all, the theoretical basis of onedimensional signal reconstruction into two-dimensional signal is described in this paper. Then, in order to make the image have the intuitionistic distinguishing feature, the Hadamard transform is introduced, which amplifies the original degree of chaos and provides the basis for qualitative and quantitative analysis. Finally, Gabor filter and Fourier spectrum analysis are used to process the Hadamard transformed image. The processed image retains some larger details such as the edge and contrast, which makes the processed image result more intuitive. The lesion lead images can be distinguished by the intuitive differentiation. At the same time, the mean value, contrast and energy entropy of the Fourier spectrum are calculated, and the results are clear. The effectiveness of the proposed method is verified by real epileptic EEG data.

This paper proposes a robust feature matching for slant SAR images based on differentially constrained RANSAC. Firstly, we obtain a set of initial matches provided by SAR-SIFT operator. Then, we set a differentially constrained model with strong constraint in the azimuth direction and no constraint in the distance direction. We use the nearest neighbor contrast (NNC) and the distance offset definition (DOD) techniques to eliminate outliers with large distance distortion. Experiments are carried out on Chinese spaceborne and airborne SAR images. The results show that the proposed method has excellent performance in accuracy, distribution and efficiency, and is suitable for matching SAR images covering areas with large terrain fluctuations such as mountainous areas.

Pedestrian fine-grained attribution detection has drawn significant interest in areas such as smart video surveillance analysis and pedestrian re-identification. Current object detection methods based on deep convolutional neural networks (CNNs) have achieved great progress. However, detections of small parts of pedestrian are still challenging due to their limited resolution and information in images. In this paper, we propose a novel CNN-based framework for improving the accuracy of small objects detection, dubbed MSF-ACNN, using atrous convolutions in cascade along with multi-scale feature fusion: 1) Atrous convolution effectively expands the field-of-view of small regions without increasing the number of parameters and computation. 2) Multi-scale feature fusion obtains more meaningful fine-grained information from both the low-level and high-level feature maps and can handle a variety of image scales. Our results show that MSF-ACNN can obtain better mean average precision (mAP) than the current state-of-the-art methods with faster detection speed, achieving significant improvements on certain small parts of pedestrian such as shoes, bag and hat.

Stereo vision is increasingly applied to various fields of life, and three-dimensional (3D) reconstruction is one of the hottest topics in stereovision. Because the active 3D system is costly and susceptible to illumination, and the reconstruction results have problems in accuracy and robustness, this paper proposes a passive 3D reconstruction method based on binocular stereo matching algorithm. Because the precision of the sparse stereo matching algorithm is not high, and most of them are applied to the indoor scenes, so the region-based dense stereo matching algorithm is adopted, which uses the correlation degree of the gray information between the windows to match, and selects the optimal similarity measure and window size by objective evaluation. After the initial disparity map is obtained, the disparity map is refined by the parallax reliability detection and the sub-pixel interpolation processing. Finally, the principle of similar triangulation is used for 3D reconstruction, and the texture mapping technology is used. The effectiveness of the algorithm is verified by experiments.

This paper presents an efficient 3D correspondence grouping algorithm for finding inliers from an initial set of feature matches. The novelty of our approach lies in the use of a combination of pair-wise and triple-wise geometric constraints to filter the outliers from the initial correspondence. The triple-wise geometric constraint is built by considering three pairs of corresponding points simultaneously. A global reference point generated according to the model shape can be mapped to the scene shape thereby form a derived point by the triple-wise geometric constraint. Then, all the initial correspondence can be filtered once via the global reference point and the derived point by using some simple and low-level geometric constraints. Afterwards, the remaining correspondences will be further filtered by means of the pair-wise geometric consistency algorithm. Finally, more accurate matching results can be obtained. The experimental results show the superior performance of our approach with respect to the noise, point density variation and partial overlap. Our algorithm strikes a good balance between accuracy and speed.

To eliminate traces in the overlap area of images caused by splicing, and reconstruct 3D face models with high-precision and photo-realism, this paper presents a multi-view texture fusion method based on triangular patches in 3D face models. The 3D point cloud is acquired according to the multi-view range images of an object and the parallax principle. By performing free registration method, ICP (Iterative Closest Point) method and a series of post-processing algorithm on the point cloud data, we can obtain a complete and non-redundant 3D geometric model with higher precision than counterpart methods. Using the brightness balance algorithm proposed in this paper, the differences of texture images in brightness and color are eliminated. The compound-weight construction algorithm combining with Gauss-Laplace Pyramid method is proposed to fulfill the 3D model texture fusion. In a variety of modeling experiments with our self-developed 3D face acquisition device, the proposed method is demonstrated to be efficient. It outperforms the counterpart methods and has practical significance in real world applications.

Dental model repair is a critical stage in the orthodontic treatment. The repaired teeth need to preserve the original morphological features, and the algorithm is required to be automatic for saving time and avoiding uncertainty caused by human handling. This paper presents an efficient and automatic algorithm to repair the tooth models' large holes. Our algorithm consists of three main steps: model's principal axis determination, model's contours processing (generation and repair), and hole surface reconstruction and fairing. We propose a novel LRBC (Linear Regression on Boundaries' Centers) algorithm to determine the model's principal axis. After generating contours based on the principal axis, we adopt point-normal cubic B-spline curve interpolation to repair each broken contour line. At last, we reconstruct holes' point-sets by constrained Delaunay triangulation, and smooth the patches with a modified Laplacian operator. Experiments on different dental models demonstrate that our algorithm can not only achieve a result with a high morphological similarity and a low error, but also satisfy the requirements of actual orthodontic applications.

This paper presents a 3D reconstruction method fusing structure from motion and laser scanning method. The precision of the 3D reconstruction is improved by combining point cloud obtained from different sensors. First, a scaled principal component analysis-iterative closest point (a scaled-PCA-ICP) algorithm is proposed to do the registration that can overcome the influence of large scale variance. Then, with the large scale factor, a recalculated registration center is proposed by region segmentation to realize the point cloud registration again. Finally, the two-view point clouds are robustly matched using the proposed optimization method to complete 3D color reconstruction of the outdoor large scenes. The proposed 3D reconstruction method is evaluated on the non-synchronous database of the real-world multi-view vision sequences obtained in an outdoor environment and laser scanner. The experiment results show improvements in both accuracy and efficiency.

Building 3D local surface feature with a local reference frame (LRF) can obtain rotational invariance and make use of 3D spatial information, thereby boosting the distinctiveness of a 3D local feature. However, this situation is based on the assumption that the LRF is stable and repeatable. Owing to the disturbances like noise, point density variation, occlusion and clutter, LRF may suffer ambiguity so that limit the ability of a LRF-based 3D local feature. This paper presents an efficient method for LRF construction. The experimental results show the superior performance of our proposed LRF in terms of repeatability and robustness on several popular datasets by comparing with the state-of-the-art methods. Moreover, our method is computational efficient as well.

Based on Delaunay triangulation, point clouds from laser-scanning, computer vision, procedural modeling are input and organized into multi-resolution alpha-shape series. An algorithm for examining singular points on the alpha-shape is proposed on the basis of the tetrahedral facet-adjacency connection from the Delaunay. A final iterative singular point reduction is performed to guarantee the 2-manifold of the alpha-shape, which generates 3D printable models.

Defect reconstruction of three-dimensional (3D) surface is a challenging task in various industry applications for its speed and accuracy. This paper puts forward a fast, robust and automatic approach which is capable of defect reconstruction by establishing a defect profile model with RANdom SAmple Consensus (RANSAC) and Iterative Closest Point (ICP) algorithm. The method mainly contains three steps: defect profile establishment, profile optimization and surface reconstruction. Employed with the 3D data obtained by a structured light scanner, we have tested on surface with different types of concave defects, the results show the effectiveness of our new technique in practical applications for defect reconstruction.

Limited-angle tomography has gained much interest in late years Nevertheless, image reconstruction from incomplete projections is a classic ill-posed issue in the field of computational imaging. In this paper, we propose a scheme based on the sparsifying operators and approximation of ℓ₀-minimization. Our framework includes two main components, one for a sparsifying operator, and one for learning the scheme parameters using ℓ₀-minimization from insufficient computed tomography data. Thus, the proposed scheme is capable of recovering high quality reconstructions at a range of angles and noise. Compared to the total-variation (TV) regularized reconstruction scheme, σ-u scheme and ATV (Anisotropic Total Variation) scheme, validations using Shepp-Logan phantom computed tomography data demonstrate the significant improvements in SNR and suppressed noise and artifacts.

Since the stencil image used for surface mount technology (SMT) always has various defects such as less holes and burrs in the laser processing and imaging, it is indispensable to detect those flaws with high accuracy. An automatic registration lies at the root of identifying defects. In this paper, a novel automatic registration algorithm for stencil images is proposed. According to the distribution probability density of the coordinates of gravity center points in a stencil image, the adaptive parameter DBSCAN clustering algorithm is adopted to classify those points. As a result, we could find corresponding gravity center points (feature points) in the stencil image and its standard design file respectively. A transformation matrix between the stencil image and its standard design file is obtained by the feature points. Experiments have shown that this automatic registration algorithm can be well adapted to the stencil images with random defects.

In this work, we address the problem of depth estimation from a single image. This is a challenging task because a single still image on its own does not give much depth cue, while recent advances in CNNs have made learning and predicting depth from a single image possible. We propose a new residual convolutional neural network (CNN) with dilated convolution and spatial pyramid pooling (SPP) structure to model the ambiguous mapping from a monocular 2D image to its depth map. The advantages of our method come from the use of dilated convolution and multi spatial scale information. Compared with existing deep CNN based methods, our method achieves much better results in indoor and outdoor scenarios.

This paper proposes a new variational model for deblurring low-resolution images, a.k.a. single image nonparametric blind super-resolution. In specific, a type of new adaptive heavy-tailed image priors are presented incorporating both the model discriminativeness and effectiveness of salient edge pursuit for accurate and reliable blur kernel estimation. With the assistance of appropriate non-blind super-resolution approaches, nonparametric blind super-resolution can be cast as a regularized functional minimization problem. An efficient numerical algorithm is derived by harnessing the alternating direction method of multipliers as well as the conjugate gradient method, with which alternatingly iterative estimations for kernel and image are finally implemented in a multi-scale manner. Numerous experiments are conducted along with comparisons made among the proposed approach and two recent state-of-the-art ones, demonstrating that the proposed approach is able to better deal with low-resolution images which are blurred by various possible kernels, e.g., Gaussianshaped kernels of varying sizes, ellipse-shaped kernels of varying orientations, curvilinear kernels of varying trajectories.

In this paper, a reversible data hiding method aims at block truncation coding compressed color images is proposed, which can reconstruct the original compressed image effectively after extracting the embedded secret data. In order to improve the compression quality, a mean absolute deviation based clustering is applied to find an approximate optimal common bitmap. The common bitmap is used to get the quantization levels of each block in each color channel. The secret data are then embedded in the quantization levels of each block by the histogram shifting method and quantization level patterns based method. Experimental results show that the proposed method outperforms the existing reversible data hiding methods for BTC-compressed color images in terms of hiding capacity and visual quality.

The aim of this study is to determine in aerial images the most significant parameters to automatic recognition of objects or textures such as house roof, vegetation, Ã Statistical method is Principal Component Analysis, a technique for classification and for examining relationships among quantitative variables. For Image Processing, Java Language is used. For statistics, SAS programs (Statistical Analysis System) with IML package (Interactive Matrix Language). Aerial Images are 400*400 pixels with 10 pixels for 1 meter on ground, with the three r,g,b components (red,green,blue). With image a first simple example and fast operator is performed to find contours, then algorithm has been adapted to dene regions with growing and merging of zones. On those zones a set of parameters is calculated to perform Principal Component Analysis for summarizing data. Results of this method allow classification of zones and find the most significant parameters to be selected, levels of r,g,b components, dispersion of those components, and one calculated value in relation with geometric characteristics of each zone.

The quality of outdoor images is significantly influenced by atmospheric environment and the vibration of imaging system. In this paper, we propose a novel model to describe the process of image degradation due to the above two factors, and a blind image restoration method based on this model. We derive the degradation model from the hypothesis that the process of atmospheric degradation and the process of vibration causing motion blur are independent. The contrast and brightness of images is decreased due to atmospheric absorption and scattering. Besides, multiple scattering and camera vibration lead to image blurring. We estimate the atmospheric light and transmittance, therefore the image restoration problem is converted to a blind deconvolution problem. A deblurring method of multi-frame image is proposed under the Bayesian posterior probability model, where the latent images are regularized by a natural image gradient prior and the point spread functions are regularized by an l₁-norm based prior. An alternating minimization approach is used to solve the optimization problem. Experimental results of synthetic images and real world images demonstrate that our model can well describe the degradation process and the proposed method can recover images effectively.

The use of convolutional neural networks (CNNs) for general no-reference image quality assessment (NR-IQA) has seen tremendous growth in the research community. Most these methods used the patches cropped from the original images for training. For these patch-based methods, the ‘ground truth’ quality of patches is essential. In practice, these methods often took the quality score of an original image directly as the labels of its patches’ quality. However, the perceptual quality of image patches generally differs from the corresponding image quality. Thus, the noise in patches’ labels may hinder effective training of the CNN. In this paper, we propose a CNN with two branches for general noreference image quality assessment. One branch of this model predicts the patch quality, and the other predicts the uncertainty, which denotes the degree of deviation of the patch quality from the image quality. Our model can be trained in an end-to-end manner by minimizing a joint loss. We tested our model on widely used image quality databases and showed that it performed better or comparable with those of state-of-the-art NR-IQA algorithms.

In this work, we propose a novel encoding-decoding based image captioning framework, which improves the performance by jointly exploring the visual object-context features, generic and specific semantic priors. In the encoding of RNN, we extract the semantic attributes, object-related and scene-related image features first, and then feed them sequentially to the encoder of RNN, which considers the rich general semantic and visual object-context representation of images. To incorporate the testing specific semantic priors in the decoding of RNN, we apply cross-modal retrieval to find the most similar captions of the testing image in the visual-semantic embedding space of VSE++. The BLEU-4 similarity is utilized to evaluate the similarity between the generated sentence and the retrieved captions, which incorporates the sentence-making priors to the testing-specific reference captions. The evaluation on benchmark dataset Microsoft COCO shows the superiority of our algorithm over the state-of-the-art approaches on standard evaluation metrics.

Hyperspectral image classification is a critical issue in hyperspectral data processing. However, the task has been acknowledged as extremely challenging due to its characteristics including high dimensionality in data, spatial variability of spectral features and scarcity of marked data. In this paper, we propose a new classification method combined with Local Binary Patterns (LBP) and Singular Value Decomposition Networks (SVDNet). Linear Prediction Error is first employed to select informative spectral bands. Then LBP is utilized to extract the texture features. After that, the extracted features of a specified field are transformed to 2-D images. Finally, SVDNet classifies the obtained images and then the classification result can be obtained. Experimental results on the real hyperspectral dataset demonstrate that the proposed method is capable to achieve higher classification accuracy or at least comparable to existing methods.

In this work, we proposed a meta-learning method for the classification of microvessel images based on PAC-Bayes. We first introduce the modeling of Single-Opponent (SO) neurons to capture the color information of microvessel images. Then, we presented the PAC-Bayes bound on multiple learning tasks for the classification of microvessel images by optimizing the PAC-Bayes objective function. Further, we summarize the meta-learning algorithm based on PACBayes to classify microvessel images in detail. The proposed method is superior in precision and f1-score compared with other representative methods.

The rapid development of meta-learning methods enables the generalized classification of histopathology images with only a handful of new training images. Meta-learning is also named as learning to learn. In this study, we propose a LSTM-model based meta-learning framework for the histopathology image classification. We apply the DoubleOpponent (DO) neurons to model the texture patterns of histopathology images. And the LSTM-model is utilized for the optimization of the meta-learning algorithm to classify the histopathology images. Experiment results on real dataset demonstrated that the proposed method leads in all the measures, namely, recall, precision, F-measure and accuracy.

With the growing mobility of the population and popularity of the Internet, real estate agents have larger database to manage. This paper presents a solution to classify images of a certain house, such as living room, kitchen, bathroom, layout sketch and external appearance collected by a real estate agent using transfer learning. The pictures are like those images posted on the real estate agent website to help people find out what’s the house looks like inside and outside. We employ a transfer learning approach for VGG-19 architecture. Using a network pre-trained on the general ImageNet dataset, we perform supervised fine-tuning on the last full connect layer and change the output size from 1000 to 5. Experimental results achieved with 5-fold cross-validation show that after training, this fine-tuning approach achieves high test accuracy of 99.4%.

Malignant melanoma (MM) of the eyelid is of high malignancy, high mortality, and easy to metastasize. Currently, the gold standard for MM treatment and prognosis is histopathology, but the diagnosis of different experts is often divergent. The computer-aided diagnosis based on deep learning helps to improve efficiency and accuracy. In this paper, a complete set of methods for MM diagnosis is proposed using the convolutional neural network (CNN) to classify the patch level pathological images. Hematoxylin and Eosin (H and E)-stained pathological images of the eyelids are classified as malignant melanoma and non-malignant melanoma (NMM). The prediction results are filled by location in the probabilistic map of the whole slide image level. Random forest classifier based on CNN inference results extract 31- dimensional features to achieve whole slide image-level classification. The color constancy method and the edge extraction mapping method based on the Sobel operator (EMBS) can significantly improve the performance of the model. The patch level classification results show that the balance accuracy is 93% on the Second Affiliated Hospital, Zhejiang University School of Medicine (ZJU-2) test set, and the balance accuracy is 89.4% on the Shanghai Ninth People’s Hospital, Shanghai JiaoTong University School of Medicine (SJTU) test set. The corresponding area under curve (AUC) is 0.990 and 0.970. For whole slide image level classification results, the AUC for SJTU test set is 0.999, the sensitivity is 100%, and the specificity is 97.4%. As a result, our model can effectively tackle the challenge of clinicopathological diagnosis and relieve the pressure of pathologists.

Coin classification automatically plays important roles in many applications, e.g., vending systems. Glossy reflection is one of the key factor that affect the performance of vision-based coin classification, especially in a complex environment. In this paper, we propose a novel method for robust coin classification. Contrary to the previous method, we get the glossy area first. Edge features and texture features are used in glossy area detection. Then the deep learning features are extracted based on non-glossy area instead of the whole coin image. Finally, the coin classification results are got from the VGG nets scheme. Comprehensive experiments show that our method is robust under various complex environments. The comparison experiments demonstrate that our method can outperform the state-of-the-art method. Our method achieves 95.80% accuracy.

Recent deep learning methods have demonstrated remarkable impact on the classification of biomedical images. In this paper, we proposed a correlation-filter enhanced meta-learning approach for the classification of biomedical images. Firstly, in the training stage, we use the training samples to optimize the model parameters of meta-learning. Secondly, in the testing stage, we utilize the data samples of the new task to generalize the model parameters. Thirdly, the nearest neighbor image from one sample batch is searched for the new instance image, with the classifying score provided by the meta-learning model. Fourthly, the template of the circular cross-correlation filter is optimized in the Fourier domain, using the new instance image and its nearest neighbor image. Fifthly, the support weight of the sample batch is calculated for the classified label by the meta-learning model. Finally, we propose the multi-batch voting mechanism to decide the label of the new instance based on the correlation-filter template. Experiments on the classification of biomedical images demonstrated the effectiveness of our approach, compared with other state-of-the-art methods.

Predicting accurate location of Protein Subcellular is conductive to acknowledging the function of protein and finding the cancer biomarkers. Unfortunately, many experimental approaches for classifying the location of protein subcellular are still high-cost and time-consuming. However, deep convolutional neural network has achieved significant advances in many fields, such as image classification, object detection and segmentation, it’s driving us to use the deep convolutional neural network to classify the protein subcellular images. Because of unavoidable differences between bioimages and natural images, for instance, the biological subcellular image texture information is not as clear as natural images. That’s means if we use a deep model to train bio-images for finishing classification task directly, its result of this experiment will be not as good as what we expected. Therefore, we utilize Partial Parameter Transfer Strategy (PPTS) and Spatial Pyramid Pooling (SPP) algorithm for achieving bio-images classification task. Using the partial parameter transfer strategy to optimize the training process of the deep model is the first step, and the second procedure is to use spatial pyramid pooling layer to optimize the architecture of deep convolutional neural network. To solve the task of bioimages classification via jointly with the above two algorithms, the performance shows that our approach can acquire better results than traditional deep learning methods.

The purpose of this paper is to determine the ripeness level and identify the variety of the Durian fruit through image processing using Principal Component Analysis (PCA) and Support Vector Machine (SVM) algorithms. The ripeness level is classified into three stages: unripe, ripe, and overripe. There are numerous varieties of durian available in the Philippines, and this study specifically uses Puyat, Arancillo, Cob, Davao Selection and UPLB Gold varieties. The study used 100 durian fruits, 80% of which were used for data training while the 20% were used for data testing. The study yields 95% overall accuracy and 5% misclassification rate.

Image salient region detection is widely used in many fields, such as automatic target detection, image retrieval, object recognition and image segmentation. Although there are many methods related to image salient region detection, improving its accuracy is still one of the hot research areas in image processing. In this paper, we propose a novel salient detection algorithm based on nonlocal filtering, which is mainly used in image denoising. The proposed algorithm stacks several similar blocks to form a three-dimensional matrix by image block matching, in which the filtered image by nonlocal filtering can better preserve image detail information than traditional Gaussian filtering. We calculate the Euclidean distance between the mean of the original image and the filtered image as the saliency map. The experimental results show that the proposed algorithm can accurately obtain the saliency map, especially, the edge of the saliency map image is much better than many existing algorithms.

In this paper, we present a new image forgery detection method via a mask filter banks which is consisted with the designed mask filters to extract the features of different channels of image and a modified a ResNet to classify the input image is tempered or not. The proposed model is proved to be capable for copy-move forgery and splicing image detection. In the mask filter layer, we first convert the image from spatial domain to frequency domain, then extract the image edge information of each channel by element-wise with the designed mask matrix. Finally, edge and noise information features of different channels were fused as feature vectors fed to a trained ResNet to do classification. Experiments on three standard datasets: the copy-move forgery image datasets MICC-F220 and MICC-F2000, splicing image manipulation datasets Columbia demonstrate that proposed method get better results than the original colour image as input method and also outperform some existing works.

Air vehicles such as aircrafts and drones have played an important role in surveillance and transportation for both civil and military applications. In this paper, we proposed a few-shot learning framework for air vehicle detection by similarity embedding, with a single moving camera mounted on another flying object. Firstly, we presented the example embedding with similarity conditioned LSTM-model for air vehicle detection. Secondly, we described the support set embedding with bidirectional LSTM-model of air vehicle training samples. Thirdly, we introduced the label prediction for air vehicle image blocks by attention kernel. Finally, we applied the fully convolutional network to segment air vehicle in the accurate bounding box. Experiment results of air vehicle detection show the effectiveness of our approach.

Aiming at automatically discovering the common objects among a group of relevant and similar images as foreground, co-saliency has become a hot topic in recent years. Previous works utilize low-rank matrix recovery on the single image, but neglect the relationship between a set of images. In this paper, we propose a novel framework to capture the coherence of common salient objects, and solve the problem when the background is clatter. The model include a novel cluster-based tree-structured sparsity-including regularization that make regions from same class have identical saliency value, and a Laplacian constraint regularization is also integrated into the model, the propose is to enlarge the gaps between common objects and background in original feature space and smooth the saliency value in same cluster. Furthermore, to facilitate the efficient, a coherence weight is identified and integrated into the model. Experiment results on three benchmark datasets demonstrate are the performance of our method compared to other stateof-the-art co-saliency models.

Achieving better performance has always been an important research target in the field of automatic image annotation. This paper draws on the current popular deep learning model for the field of automatic image annotation. We propose a multiple convolutional neural networks (CNN) combination model for image annotation, which achieves satisfactory performance. First of all, we use three classical convolutional neural networks, and subsequently we examine the annotation accuracy for each CNN model. Then we take full advantage of the powerful feature representation capabilities of deep CNN, thus the last two layers of the deep CNN are extracted for each model and merged to form a new combined feature. Finally, we form our combination models by concatenating these features from each CNN model, and utilize these concatenated features to linear SVM classifier for image annotation. Experimental results on ImageCLEF2012 image annotation dataset illustrate that our combination method outperforms the traditional classifiers and the individual CNN models.

In this work, we present a novel network named CRRCNN (Cascade Rotational Region-based CNN) to detect dense objects with oriented bounding boxes. The CRRCNN consists of a Faster RCNN and a Cascade RCNN with Rotational RoIAlign. The Faster RCNN consists of RPN (Region Proposal Network) and RCNN (Region-based CNN). RPN generates horizontal bounding boxes. Rotational region proposals are generated through quadrilateral vertices regression of RCNN, and therefore Faster RCNN is regarded as a Rotational Region Proposal Network (RRPN). To generate accurate rotational bounding boxes, a Cascade RCNN with Rotational RoIAlign is proposed following the Faster RCNN, which will be demonstrated to be crucial for accurate arbitrary-oriented object detection, especially for dense objects. Feature Pyramid Network is also employed to obtain rich context information. The two networks mentioned above are unified and learned end-to-end by jointly optimizing. Experiments on the challenging DOTA dataset demonstrate the effectiveness of our approach.

Complex manufacturing makes the screens one or more defects simultaneously. Effective detection of defects is crucial in the manufacturing processes of mobile phone glass screens. In this paper, we concentrate on extracting and classifying some typical defects, such as scratch, bruise, pit and blister. Firstly, we use morphological filter to smooth background noise, and an improved gamma grayscale transformation is proposed to enhance the contrast. Then, a double threshold segmentation algorithm, based on area threshold and gray threshold, is presented to extract defects from background. Finally, according to different optimal segmentation feature values of different defects, a binary tree classifier is constructed to classify defects. The experiment results show that the proposed method can extract and recognize typical defects precisely.

Saliency detection is a fundamental problem in computer vision and image processing. Deep learning models perform better than traditional approaches but suffer from their huge parameters and slow speeds. To handle these problems, in this paper we propose the Dense Dilation Network(DDN) -- a novel saliency model based on Dense Convolutional Network(DenseNet), all layers' feature maps are extracted in the same resolution through a step-wise upsampling procedure which contains dilated convolution filters and deconvolution filters. To derive saliency maps, we propose a fusion dense block which contains dilated convolution filters and 1×1 Conv layers to fuse all low-level and high-level feature maps. DDN is an end-to-end architecture which does not need any post-processing procedures. Experiments on the benchmark datasets demonstrate that DDN outperforms the state-of-the-art methods on saliency detection and it requires less parameters and computation time.

This paper presents an innovative method to align two handwritten Chinese characters modeled by three-dimensional point sets. A piecewise cubic Bézier curve is first defined for each character stroke, interpolating the 2D sample points. The density of sample points is then adjusted along the curve for even distribution. Later all 2D points are extended into 3D space to integrate additional geometry information. Registration is then performed between two 3D point sets using Gaussian Mixture Model (GMM). This paper establishes the registration by minimizing the Euclidean Distance(L2) between two GMMs and then apply transformation accordingly. The key idea of this process is to include extra geometric information of the calligraphy to facilitate the registration process, using simple algorithms. Through the alignment and transformation process, the global layout of character is adjusted, but with local features retained. This technique can be applied in the morphing and beautification of Chinese calligraphy.

In this paper, an advanced method is proposed concerned with the problem of classifying objects laying on the seabottom. This method is a fusion of different methods (KNN, SVM and improved CNN) mainly consisting of three steps: Firstly, acoustic images of underwater objects are pre-processed and segmented into shadow and sea-bottom expressed as binary images. Then Zernike moments of these binary images are computed as feature vectors and they are classified by a k-nearest neighbor (KNN) classifier. At the same time, a support vector machine(SVM) and an improved convolutional neural network (CNN) are also used to classify binary images. Finally, a vote classifier combines classification result of the three classifiers (KNN, SVM and improved CNN) and give the result. The method is applied to synthetic aperture sonar(SAS) datasets for validation. Comparing with each individual classifier (KNN, SVM and improved CNN), the proposed method performs stable and achieves better accuracy.

This paper proposes a digital image encryption algorithm based on dynamic DNA encoding and operations using hyper digital chaos in frequency-domain. The chaotic DNA encoding and operations are applied to diffuse and scramble both of the amplitude and phase components in frequency-domain. The proposed multi-fold encryption algorithm is evaluated via numerical verification, in which the ideal performances of the encrypted images are achieved, in terms of uniformity in histogram, entropy as well as correlation. The results show that, since the chaotic diffusion and scrambling are introduced in the proposed encryption, each cipher-pixel is influenced by all of the plain and cipher-pixels, which improves the plain-text sensitivity of the encrypted image, and the secure performances against the differential attacks. Moreover, high security level of the encrypted image is achieved with a huge key space, due to the extremely high sensitivity of hyper digital chaos applied in the proposed encryption scheme, therefore it has the strong capability to resist brute-force attacks. In addition, the statistical characteristic of the original image is destroyed completely, thus it improves the security against the statistical attacks.

The performance of computer-vision based face image retrieval system declines significantly when large illumination, pose, and facial expression variations are presented. To tackle such a problem, we propose a closed loop face image retrieval system with implicit eye-tracking based feedback. It combines the state-of-the-art computer vision method Face++ with the powerful cognitive ability of human. In this system, the Face++ provides initial retrieving results corresponding to a target sample face image whose top ranked 36 images are then displayed on the screen for collecting eye-tracking data of the users. Upon mining the user’s cognition results from the eye-tracking data with a deep neural network and feeding them back to the system, the system begins its new round retrieving process. Experimental results from 10 volunteers in a face database containing 1,500 images of 50 celebrities show that the performance of our system becomes better and better over iterations and finally our system achieve an average precision of higher than 0.918 and an average recall rate of higher than 0.897 upon convergence.

Watermarking is widely used for copyright protection and authentic use of patented objects. Digital objects such as multimedia images, videos and data are watermarked with originator logo or the personal information of the purchaser or the both to issue authentic right to use such objects. The watermark embedded in the object changes the object quality to various amount depending on the technique used. We propose a watermarking technique in which the object remains unchanged, and hence called clean. The watermark is embedded multiple times in different pixel sequences of the object. The pixel sequences are generated using chaotic logistic maps. Multiple secret information sequences are derived after multiple embedding, which may be utilized against adversarial attacks for claiming rights or unauthorized usages.

This paper describes an new method for online handwriting signatures verification. The algorithm is based on "Siamese" deep neural network. This network consists of two identical sub-networks joined at their outputs. During verification the two sub-networks extract features from two signatures, while the joining fully-connected network measures the distance between the two feature vectors to determine whether the signature is genuine. The most remarkable advantage of the system is that it can be trained end-to-end without any handcraft feature extraction except some necessary preprocessing. Experiments on the publicly dataset yielded the performance of 4.5% equal error rate (ERR).

Image storage and retrieval incur high storage and computation complexities and cloud computing offers an attractive choice for the image storage outsourcing and the on-demand access. Since the outsourced images often include sensitive information, privacy-preserving image outsourcing scheme becomes the biggest research hot spot. In this paper, we proposed a lightweight-client and privacy-preserving content-based image retrieval (CBIR) scheme, which has the following advantages: (1) to improve search accuracy, color and texture feature are integrated to construct index; (2) to reduce the client (owner and user) burden, computationally intensive tasks, texture feature extraction and index construction, are securely outsourced to cloud server; (3) to enhance search efficiency, a two-layer index structure is also adopted. The security analysis and the experiments show the security and efficiency of the proposed scheme.

This paper presents a new approach to texture image retrieval based on the statistical modeling in the multidirectional complex transform domain for effectively improving the performance of the retrieval system. Firstly the basic directional filters are selected and used for the complex transform. Then the coefficients of highpass subband are modeled as generalized Gaussian distribution; the magnitude coefficients of complex directional bandpass subbands are modeled as Gamma distribution; and their distribution parameters are efficiently combined with the statistical features of relative phases of complex bandpass subband coefficients. Finally the texture images of VisTex standard database are retrieved using a new combined form of similarity measurements. The experimental results show that, in comparison with the existing texture image retrieval methods, the proposed approach obtains a superior average retrieval rate.

This paper proposes a novel two-tiered retrieval method for CAD models based on the graph spectral theory and face code to match three dimension (3D) CAD models with undirected graphs. This approach applies a hierarchical feature descriptor, a two-tiered retrieval mechanism to represent and retrieval CAD models. The first tier is rapid rough matching, which is a similarity metric by weighted bipartite graph matching between library models and retrieval models. Second tier is exact matching through comparing face code. Experimental results show the proposed method could improve the efficiency of retrieval performance, and achieve the reuse of CAD model.

Sketch-based image retrieval (SBIR) technique has progressed by deep learning to learn cross-modal distance metrics that relate sketches and photos from a large number of sketch-photo pairs. However, datasets of sketch-photo pairs are small, as acquisition of a large number of such pairs is expensive. To alleviate the issue, data augmentation via image transformation such as scaling, flipping, rotation, and deformation has been widely adopted. Still, insufficiency in training set seems to have impeded deep learning from achieving its full potential for SBIR. In this paper, we propose a novel data augmentation approach dedicated for SBIR. A deep neural network called Photo2Sketch (P2S) converts photos into line drawings that are visually similar to those sketched by human. An artificially augmented training dataset of sketch-photo pairs is generated at low cost by feeding photos from a large image corpus into the P2S. Experiments evaluate quality of sketch-like images generated by the P2S as well as efficacy of the proposed data augmentation algorithm under SBIR scenario. In particular, retrieval accuracy is significantly improved when the proposed algorithm is combined with the data augmentation by image transformation

Digital watermarking, as an important information security technology, can play a dual role of evidence tracing and encryption in Mobile Police System. In the process of uploading photographic pictures, the poor connectivity and stability of wireless network will greatly affect the robustness of watermarking algorithm. In this study, we proposed a blind watermarking algorithm based on SIFT feature points meanwhile with the idea of spread spectrum.The algorithm extracts sufficient sub-blocks of the host image using feature points to achieve high invisibility and extends multiple times of the original watermark to achieves high robustness and security. Experiment results show that the invisibility and robustness of watermark by using SIFT feature points are better than several other feature extractors and the algorithm meets the application requirements of Mobile Police System.

For personal-oriental applications, the outstanding recognition ability has caused great concern for privacy disclosure. Existing methods mainly protect the privacy by adding occlusion or noise to the original images before uploading, which is either impractical or destroying the utility of the photos. In this paper, we propose a Generative Adversarial Network (GAN) based image privacy protection algorithm PriGAN, which generates a privacy image for each original image to fool the recognition networks. When generating the privacy image, we consider the technique of adversarial image perturbations (AIP), which could confuse recognition networks with slight perturbations. That is, the privacy image could protect the privacy information by confusing the neural network hosted by the service providers. Meanwhile, the privacy image appears unmodified compared to the original one for human observers, and thus its utility could be preserved. The advantage of PriGAN is that it provides a general privacy protection framework for personal applications, by which privacy information of images can be protected without destroying the image utility. The experiment shows that the privacy images are misclassified with approximate 82.9% higher than the original images which indicate that our approach can prevent the privacy from leakage with considerable improvement.

Magnetic resonance imaging (MRI) plays an important role in breast cancer detection and diagnosis. Breast region segmentation on MRI images is an essential step for many analysis tasks such as the assessment of background parenchymal enhancement (BPE), the analysis of non-mass enhancement (NME) and investigation of the tumor characteristics. However, automated breast segmentation remains a challenge because the noisy background and the blurred boundary between breast and pectoral muscle on MRI images. In this study, we developed a coordinate-guided U-Net (CU-Net) to identify the boundary of the breast on MRI images. The CU-Net applied the breast position information to the segmentation of breast by adding the X coordinate channel and Y coordinate channel to the input image. A dataset of 80 3D bilateral breast MRI scans including 64 cases in training/validation set and 16 cases in test set were collected. All the input images were resized to 256×256 pixels and the intensity are cropped and rescaled according to the percentile of cumulative histogram to normalize the image. To add X and Y coordinate channels, the middle point of the breast-air boundary was detected as the origin of coordinates. The position of each pixel in the coordinates was then encoded into the input of the network. The CU-Net was evaluated with three measures: the dice coefficient (Dice), the intersection over union (IoU) and the hausdorff distance (Hdist). The results of CU-Net were compared to those obtained with U-Net method. In the test set including 16 cases, the mean Dice, the mean IoU, and the mean Hdist of CUNet were 94.1±1.1%, 92.5±1.3%, and 6.17±1.4 mm, while the corresponding measures by U-Net were 91.4±2.4%, 89.7±1.7%, and 6.98±1.5 mm, respectively. The results showed that our CU-Net was able to segment breast on MRI images reliably and was more accurate comparing to the U-Net.

Segmentation of constituent shapes is a vital yet challenging step in any automated sketch analysis and interpretation system. Conventional stroke level sketch segmentation techniques perform well on a single shape, nevertheless, their performance degrades in a cluttered multi-object sample. On the contrary, object level techniques rely on proximity based perceptual grouping for shapes with disjoint constituent parts, which requires high level semantic knowledge. To overcome these challenges, we propose the use of state-of-the-art convolutional object detectors for the detection and segmentation of hand drawn shapes from offline samples of a neuropsychological drawing test i.e. Bender Gestalt Test (BGT). Experiments with different combinations of convolutional meta-architectures and feature extractors show that such networks can successfully be employed for sketch segmentation purposes even with limited training data and resources. Amongst all network combinations under evaluation in this study, Faster R-CNNs with ResNet-101 as feature extractor, outperform others by achieving precision, recall and F-measure values of 92.93%, 95.24% and 94.07% respectively.

This paper presents a new stereo based on effective cost aggregation. Firstly, a color-based image segmentation algorithm is used to segment the reference image into blocks with different colors, and they are taken as the matching window to search the similar region in the target image. Secondly, in order to eliminate the mismatching of brightness difference and noise, Census transform is conducted with both reference image and target image to obtain the bit strings. At the same time, for the sake of matching efficiency, it is of great necessary to rectify the dynamic disparity scopes of the pixel to be matched by the disparity of neighborhood pixels. At last, the bit strings can be used to execute the cost calculation by a new kind of cost aggregation function. By moving the window in the horizontal direction, we can figure out the percentage of the mismatch pixels in every segmented block respectively. The minimum value means which corresponding the horizontal shift is expressed as the best disparity of this segmented block. The experimental results demonstrate that our method can not only improve the accuracy of the depth discontinuities, but also has the high calculation efficiency. Besides, it has good robustness in different brightness environments.

The main contribution of this work is the proposal of a densely connected convolutional network for semantic segmentation, which strengthens utilization of features and improves segmentation results even with limited training samples. To achieve this, we combine the U-Net network and our resulting system is called Dense-U-Net. Compared to traditional convolutional networks such as U-Net, there are additional concatenation layers between each pair of convolutional layers which have the same size of outputs in our Dense-U-Net, each layer can get the feature-maps of all its preceding layers as inputs while its feature-maps can be passed to all subsequent layers, and a higher segmentation quality can be achieved without a need for increasing the volume of datasets finally. We evaluate our proposed architecture by segmentation accuracy, foreground-restricted rand scoring after border thinning V^Rand and foreground-restricted information theoretic scoring after border thinning V^Info at the same time, and the results are shown on three different segmentation tasks: ISBI challenge 2012 for segmentation of neuronal structures in electron microscopic stacks, ISBI cell tracking challenge 2014(Glioblastoma-astrocytoma U373 cells) and 2015(HeLa cells), our Dense-U-Net achieves better results than U-Net and several other state-of-the-art networks on all tasks.

Breast tumor segmentation is an essential step in a computer aided diagnosis (CAD) system. Due to the speckle noise, various size and locations of lesions, it is a challenging task to segment the tumors on breast ultrasound (BUS) image accurately. In this paper, we propose a region growing method which applies the similarity score and homogeneity based on the neutrosophic set (NS) domain to segment and detect tumors. At first, the original BUS image is transferred to the NS domain and calculate the different NS elements to obtain the NS image. Then, a thresholding method and morphology method are used to locate seed regions, and after that each seed region grows separately. The direction of region growing depends on the homogeneity vector, similarity set score vector and distance vector between the candidate points and seed regions. The suspensive condition also lies on the homogeneity vector and similarity set score vector between the seed region and the NS image. Finally, a pretrained deep learning network and transfer learning scheme are used for false positive reduction. Experiment on various clinic BUS images has suggested that the proposed method is able to segment the BUS image and extract tumor accurately.

Pointing precision of attitude stabilization platform is important for airship during its flight and it is one of the key factors to make sure that the antenna isolates disturbance from airship and tracks target with stable attitude. Therefore, in this paper, a method of real-timely measuring pointing precision of attitude stabilization platform based on image processing is established to test attitude stabilization platform performance. In addition, a method of real-time extraction for target with Harris algorithm based on local Otsu threshold segmentation is established for real-time automatic measurement. This method overcomes the problems of long shooting distance, small imaging area of target and the bad light. The experimental results show our method can real-time measure the pointing precision for attitude stabilization platform and the measured value fits well with the design value .

This paper explores a Riemannian steepest method for fast converging local binary fitting model. The proposed method takes advantage of intensity information in local regions, which solves the intensity inhomogeneous images with satisfactory results. Furthermore, the Riemannian steepest descent method can be employed to local binary fitting model from exponential family and achieves convergence fast. The main contribution of this paper is that presents a general closed-form expression for the manifold’s Riemannian metric tensor of local binary fitting model, which makes the computation of Riemannian gradient flow possible. In addition, to ensure the accuracy of the segmentation results, we regularize the level set function by Gaussian smooth operator. Experimental results for synthetic and real-life images show satisfactory performances of proposed method.

In this paper, we propose a model based on DenseNets and scale invariant feature transform (SIFT) keypoint match technology for object detection and segmentation. DenseNets is built on Convolutional Neural Networks (CNNs) with dense connections and used for semantic image segmentation. Our main idea is that, on the basis of the DenseNets model, we conduct the morphological processing, and apply the SIFT keypoint match technology to detect the object pixels. Opening operation and closing operation are the basic operations of morphological processing. They all consist of erosion operation and dilation operation but the order is different between them. The morphological processing combines two kinds of operations and can form a morphological filter which can filter the noise. The SIFT keypoint match algorithm is widely used to extract the invariant of position, scale and rotation so we use it to eliminate the misjudgment. Our experiments show that our method can acquire more accurate results compared with DenseNets model.

The study of plants in space plays an important role in serving astronauts. Automatic segmentation of space plant image (SPI) provides an effective method for studying plants, and many plant segmentation methods have been proposed. However, segmentation of SPI is still challenging. Because the number of SPI is small, which greatly increases the difficulty in model training (especially deep-network-based model). For dealing with this problem, we propose a plant segmentation method based on a generative adversarial network. Our method consists of a generative network (GN) and a discriminant network (DN). The GN firstly extracts features from an input image, and then generates a feature map by developing multiple convolution and deconvolution layers. The DN merges the feature map with an actual plant image, and then computes a segmentation result by a deep convolutional network. In DN, the addition of the feature map improves segmentation accuracy of DN, and reduces the requirements of training images during training DN. Several experiments are made, and the experimental results show that our method performs well when a small number of training images is provided for model training.

Lane detection is the core module of autonomous cars and plays an important role in lane keeping and trajectory planning. Traditional works on lane detection depend on hand-crafted features and are weak to road scene variations. Recent end-to-end approaches leverage deep learning models, but are computationally demanding and need heavy work of labeling lanes. In this paper, we combine the advantage of traditional method and deep learning based method. Our lane detection method has three key novelties: (1) we use semantic segmentation based deep learning to extract region of interest (ROI), which can free the work of manually selecting features and fast locate the lane area. (2) We propose a lane feature points extracting algorithm based hierarchical clustering to effectively remove the disturbance of the noise. (3) we make use of the similarity of the inter-frame to correct the lane fitting results. Experimental results show that our lane detection method can adapt to various road scene and significantly decrease the false positive rate.

The presence of speckle noise seriously affects the application of synthetic aperture radar (SAR) images in image fusion, so it is especially important to suppress speckle noise. According to the formation mechanism of speckle noise, this paper proposes a SAR image speckle noise removal algorithm based on improved anisotropic diffusion. The algorithm improves the diffusion coefficient c(x) based on the P-M equation diffusion filter algorithm, and adds the iterative termination condition, and obtains the filtering algorithm suitable for SAR images. This method can not only solve the problem that there are many isolated noise points in the traditional P-M model filtering, but also has a good effect on image edge preservation. The simulation results show that the improved P-M model can eliminate noise well and maintain the edge information of the image well.

Aerial image stitching is a significant technology in the field of UAV applications. The state-of-art methods for aerial image stitching are always difficult to simultaneously guarantee the accuracy and efficiency. Aiming at the bottleneck, a new aerial image stitching algorithm based on group control is proposed. It mainly provides a hierarchical aerial image stitching framework, which is constructed by a dichotomous grouping and stitching process. Meanwhile, the direction of stitching is also controlled to prevent error accumulation. Additionally, the time consuming for stitching can be further cut down without losing accuracy, which attributes the success to the multithreaded concurrent implementation. Experimental results validate that a more attractive aerial image stitching performance can be achieved with the proposed algorithm.

Alternating proximal algorithm is presented for L₁/TV^p (0 < p < 1) nonconvex variational model, which typically outperforms popular models with convex variational models in restoring sparse images with corruption by the impulse noise or other outliers. Numerical experiments are reported to illustrate the effectiveness of this algorithm.

Underwater image generally suffers from color degradation and poor visibility due to the attenuation and scattering of the propagated light. To restore the real environment of the underwater, most of previous methods only consider the scattering in the polarimetric image recovery model. To compensate for the absorption in different wavelengths, a simple but effective underwater image enhancement method that based on the adaptive polarimetric image recovery model is proposed in this paper. In order to achieve automatic estimation of model parameters, the backscattered light estimation is based on the quad-tree decomposition to extract the fully backscattered area automatically. The experimental results demonstrate that the visual quality of underwater images are efficiently enhanced.

This paper proposes a method for image smoothing which is invariant under general coordinate transformation. This method is based on a position dependent metric tensor which transforms appropriately against the general coordinate transformation. Using this metric tensor, a method for invariant image smoothing against the general coordinate transformation is constructed. Effectiveness of the proposed method is confirmed by computer experiments.

Due to the lack of color and “halo artifacts” in enhancement of low illuminance images based on existing algorithms, this paper presents an image enhancement algorithm based on an improved Lab color space Retinex algorithm. The author innovates the image preprocessing, improves the guided filtering considering the deviation of the guided filtering algorithm, and proposes a new method to enhance the filtered image. The overall idea of image enhancement in this paper is as follows. First of all, the low illumination image is preprocessed so that the obvious dark regions in the image are roughly enhanced and adjusted. Then the preprocessed image is converted to Lab color space and the luminance component (L) and chroma components (a, b) are separated out. Using a modified guided filtering algorithm to estimate the incident component luminance value to obtain the reflected component, after the gamma correction is performed, the image brightness contrast is improved by the adaptive logarithm mapping method. Finally, the brightness component and chromaticity component are recombined and converted to RGB color space. Experiments show that the improved algorithm proposed in this paper has a better image enhancement effect and can solve the "halo artifact" problem of the image.

In order to better suit the continuous development of medical environment and protect the medical image security, a novel region of interest (ROI) of medical image encryption scheme based on two-dimension discrete wavelet transform (2-D DWT) and five-dimension hyperchaotic system is proposed. In this paper, the medical image is first separated into ROI and background area, and one low frequency and three high frequency components in the horizontal(H), vertical(V) and diagonal(D) directions of ROI are obtained via 2-D DWT. Next, only the low frequency component which approximate the original image is scrambled and spread by chaotic sequence. Then, the reconstruction ROI of image is generated by using encrypted low frequency and components of H, V and D. Fin1 ally, the confusion-diffusion process is used to encrypt the reconstruction ROI of image. Our scheme can realize the encryption of medical image not only in space domain but also in frequency domain. Experiments simulation results, such as key sensitivity tests, correlation analysis, information entropy tests, known/chosen plaintext attack analysis and comparison tests, show the advantages of our scheme.

Tensor has been widely used in computer vision due to its ability to maintain spatial structure information. Owning to the well-balanced unfolding matrices, the recently proposed tensor train (TT) decomposition can make full use of information from tensors. Thereby, tensor train representation has a better performance in many fields compared to traditional methods of tensor decomposition. Inspired by the success of tensor train, in this paper, we firstly apply lowrank tensor train to recovering noisy color images. Meanwhile, we propose a novel algorithm for noise-contaminated images based on the block coordinate descent (BCD) method. The numerical experiments demonstrate that our algorithm can give a better result in the real color image both visually and numerically.

This paper describes a novel method for correcting a color cast in an image which contains RGB and near-infrared crosstalk. We name this type of images as four-band images. The algorithm is based on the back propagation neural network. When a four-band image is put into the algorithm, it will use a trained weight matrix to transform this image into the normal one. This matrix is obtained by a BP algorithm which is used to find the implicit relationship between the input image and the corresponding target image. The size of the matrix is determined by the number of hidden layers of the neural network and the number of rows of the pixel matrix of the input and output images. In this paper, we acquire the weight matrix of the input layer to the hidden layer with a size of 506*486 and the weight matrix of the hidden layer to the output layer of 486*506. The camera used herein is a slightly modified CMOS sensor that replaces the IR-Cut filter with an 850 nm bimodal filter. The dataset used in this paper has five pairs of one-to-one correspondence images. One of them is a three-band image with normal color using the original IR-Cut filter, and the other is a four-band image taken by the modified CMOS sensor. The proposed method has been tuned and tested with positive results in this dataset.

An eye state analysis (i.e., open or closed) is an important step in the detection of fatigue driving. In this paper, a weighted color difference matrix algorithm is proposed for analyzing a driver’s eye state. First, an image of the driver’s eye is obtained from a face detection database. Two feature images are then constructed for the eye image, which are each gray-scale normalized. The feature image is then projected into a block feature matrix, and the feature value is calculated to construct a feature vector. Finally, a support vector machine is used to train and classify the extracted eigenvectors, and the state of the driver’s eye is judged to further analyze the driver’s state of fatigue. To evaluate the performance of the proposed algorithm, experiments are conducted with several publicly available databases, showing that our algorithm is efficient and reasonable.

Breast cancer has become a worldwide disease in recent years. However, despite its growing prominence, the number of pathologists equipped to handle these cases is insufficient. Computer-aided diagnosis (CAD) system contributes to reduce costs and improve efficiency of this process. A framework based on convolutional neural networks (CNNs) which could be used to automatically detect the multi-class cancer areas on gigapixel pathology slide images was proposed. Moreover, combining the slide image characters, rescale and careful data augmentation methods were used to train the patch-based model with a small dataset. To validate the developed framework, we conducted experiments with Breast Cancer Histology Challenge (BACH) dataset and obtained International Conference on Image Analysis and Recognition (ICIAR) score of 0.582, outperforming the second-place finisher in BACH2018, for the 4-class tissue segmentation task.

Medical imaging quality is limited by physical, technology and economical considerations. In the practical application, bilinear or bicubic interpolation are performed to enhance the resolution of medical image. However, these interpolation methods will lead to undesirable jagged artifacts and blur details, which would affect the performance of post-processing such as segmentation and registration. Example learning based single image super-resolution (SR) is an economical way to recover the high-resolution (HR) images from a low-resolution (LR) image. The learning based methods are to add the high frequency information borrowed from the learning data to the input image by referencing the similarities between the input data and the learning examples. Due to the incompatibility between the examples and the testing image, this paper proposes a novel multiscale rotational similarity based single image SR method, in which the learning examples are generated from the LR image itself. The motivation of the proposed method is that the small patches in medical image tends to repeat themselves across different scale or different angle. With the intensive investigation of the similarity redundancy in the multiscale and multiangular levels, the relationship between the patches of HR image and that of LR images will be estimated. Subsequently, the HR images will be generated from a single LR image in a gradual magnification framework. Meanwhile, a nonlocal prior regularization terms will be utilized to enhance the accuracy of reconstruction. Some experiments are performed to indicate the feasibility and effectiveness of the proposed algorithm.

Vaginitis, the most common disease of female genital tract infections, mainly relies on the morphological detection of the vaginal micro-ecological system to diagnose under the microscope. It affects women's normal life seriously, even their fertility. Since the morphological detection is very dependent on the experience of the observer, while the experienced doctors are mostly concentrated in large cities, the problem of diagnosis of vaginitis in rural women is extremely serious. Convolutional neural network (CNN), the typical algorithm of artificial intelligence, has shown great potential in many visual classification tasks. However, it is difficult to apply CNN method directly to the diagnosis of vaginitis. To solve the problem, this paper proposes an algorithm combining CNN with decision-making tree (CNNDMT) based on medical expert consensus. In a way of incorporating features automatically extracted by the machine and expert knowledge, automatic diagnosis of vaginitis disease is realized. Experimental results show that the CNN-DMT approach improves test accuracy by 8.46% over the leading CNN method, while enhancing the accuracy of normal bacterial flora by more than 15%.

Glaucoma is among the leading causes of irreversible blindness worldwide. Early detection and proper treatment can help delay disease progression and in turn associated visual impairments. Glaucoma is commonly linked with elevated eye pressure leading to damaging of the optic nerve. Retinal images have been successfully utilized for glaucoma diagnosis by extracting relevant features related to the optic disc, most commonly from the RGB color model. In this study, the relevance of statistical features computed from the detected optic disc region are compared for the RGB, HSV, and CIEL* a * b* color models. Feature computations required less than 0.2 seconds per image for the complete feature vector. Correlation feature selection was employed to determine the most relevant channels and associated features in each color model. Analysis showed that the red, value, and b* channels of the RGB, HSV, and CIEL* a * b* models, respectively were the least useful for glaucoma detection. Furthermore, statistical features computed from the HSV model were found to result in superior results in comparison to statistical features extracted from the RGB and CIEL* a * b* models. The HSV statistical features implemented in this work were able to classify retinal images into normal or glaucomatous with accuracies of 90.9% and 92.4% along with areas under the receiver operating curve (AUC) of 0.953 and 0.944 using the support machine vector (SVM) and naïve Bayes classifiers, respectively.

Esophageal cancer is one of the leading causes of mortality worldwide. Early diagnostics are imperative of improving the chances of correct treatment and survival. Pathology examination requires time consuming scanning while often leads to a disagreement among pathologists. The computer-aided diagnosis systems are critical to improve the diagnostic efficiency and reduce the subjectivity and error of human. In this paper, a deep learning–based approach is proposed to classify the H&E stained histopathological images of esophageal cancer. The histopathological images are firstly normalized to correct the color variations during slide preparation. Then one image patch is cut into five pieces and gets five corresponded features via convolutional neural networks (CNNs). Subsequently, the five features of one image patch are fed into the Long Short-Term Memory (LSTM) model for further feature extraction and integrated as one for the last classification. The experiment results have demonstrated the proposed approach is effective to classify the esophageal histopathological images with the accuracy of 84.5% which outperforming 10 percent than GoogleNet.

Produce orthopedics surgical guides with the assistance of computer-aided design and three-dimensional (3D) printing techniques, to provide supports for doctors before or during surgeries. We develop a software which was programmed by C/C++ language, using the development tool MFC for interface and OpenGL for graphics processing. The software was used to personalized orthopedics surgical guides, and the finite element analysis of the orthopedics surgical guide was carried out. And then the solid model of the orthopedics surgical guide was printed through 3D printer. The finite element analysis of the 3D model shows that the model can meet the requirements in accuracy and stability and the final generated model of the orthopedics surgical guide can closely fit the spine bone model.

Accurately identifying the location of hard exudate in retinal fundus helps to reduce the risk of blindness caused by diabetic retinopathy. An approach based on IHBMO-RF was proposed for hard exudate detection. The preprocessing procedure contains brightness correction, nonlinear wavelet denoising and high-hat filter enhancement. We used 2D Gaussian matched filter to locate the center of the optic disc and eliminate it. Secondly, the morphological methods were applied to obtain exudate candidate areas. Finally, the candidate regions were classified by using the IHBMO-RF algorithm to extract the final and accurate hard exudate. This method was tested in the published DiaretDB1 fundus image database and the accuracy of extraction of hard exudate was 96.8%. The experiment shows that based on the IHBMO-RF algorithm can accurately identify the position of hard exudate in fundus images and has good robust performance.

A calibration device for the optical scanner is presented in this paper for the situation that lacking of a convenient calibration method for the optical scanner in domestic industry practice. The device utilized the spherical center distance of fixed balls as standard dimensions. It can provide various dimensions in different orientations and magnitudes since it has nine fixed standard balls in different positions. When we run an actual measurement process on this device with an optical scanner, the indication error of the scanner can be obtained conveniently. The device is simple in structure and easy to operate, it can produce multiple dates in one single measurement process simultaneously, which can evaluate the performance of the optical scanner comprehensively.

This research presents the character input system by using wearable camera based gaze estimation system (GES) for disorders. The GES is the system that can estimate the user's gaze direction. The proposed system uses the inside-out camera to get the user's gaze image. This system uses the gaze image to classify the character and create the word. Proposed system consists of two processes: Gazes estimation process and character recognition process. This research uses the Convolutional Neural Network (CNN) method to archive the good accuracy. By using the proposed method to train the model, the system was confirmed through evaluation experiments. The experimental results show the good accuracy of the proposed system compared with the previous research.

Light field camera has the ability to capture ray’s position and direction information through a single shot. The multiview images generated from the light field image provides unique advantage for depth estimation. In this paper, a method for 4D light field depth estimation was proposed by exploiting a hybrid stereo matching method with multidirectional multi-baseline stereo matching and K-Means segmentation. The multi-baseline stereo matching is adopted to effectively utilize the information among multiple low resolution sub-images. The multi-directional consistency checking is used to identify the fallible matching results due to occlusion or lack of texture. The K-Means segmentation method is used for locate the object boundary and hence ensure that the match window lies within a same object. A weighted median filtering method is adopted to further increase the quality of the disparity map. The experiment verifies the effectiveness of the method used. We evaluated the performance of our method on a publicly available dataset. Result shows that the proposed method can achieve a competitive result with a reasonable running time.

In range test, for measuring the target attitude in target terminal trajectory, we developed a fast Hough transform to real-timely extract axes of the axes symmetry target in high-speed image processing system. The theoretical analysis shows that the accuracy of this scheme is higher than that of standard Hough transform, and the closer to one the absolute value of slope is, the lower the error is. Experimental result shows that the computing speed for line detection of this scheme is higher than standard Hough transform, and the greater the noise is, the more obvious the advantage of speed is. In addition, the error distribution of the experiment is in agreement with the theoretical error distribution.

As an important technology in the film industry, Multispectral Lighting Reproduction is used widely to provide real and vivid images for audiences. The overlapping shadows cause the redundant part of the film images which lead to uncoordinated visual effects. This problem is confusing the film industry. To avoid this problem, light sources have to irradiate from a far distance inside the gigantic building, meanwhile a high cost of equipment is required. Hence, it is important to propose a better method which can shorten the illuminative distances and save a huge amount of cost while maintaining the good accuracy of lighting reproduction. In this paper, we propose a novel system to solve the overlapping shadows problem, which can remarkably reduce the power and hardware costs in multispectral lighting reproduction system. The system consists of two workflows. First, the non-negative least squares solution is used to solve the illuminance coefficient. The accurate light is achieved by the illuminance coefficient. Second, images are categorized as the full shadow, overlapping shadows and non-shadow by different luminance and color features. Finally, overlapping shadows are removed by the inpainting algorithm. Experimental results show that our method can effectively detect the overlapping shadows and images inpainting of the dataset.

Facial illumination severely affects the face recognition performance; thus, it should be finely compensated beforehand. This paper mainly studies three representative illumination-insensitive representation methods including GRF, WF and LBP, all of which consider using the local pattern information to extract facial features. We first present the ideas and theories of each method. Then based on the reflectance model, the underlying connections of GRF and WF are discussed through showing the deduction of GRF and WF how they exclude the luminance component and are only related to the intrinsic facial features. We also give the explanation about the correspondence of LBP to the reflectance model. Finally, experiments on a standard but challenging illuminated face database are conducted, in which GRF, WF and LBP are tested and compared to other illumination normalization methods in terms of face recognition rate.

Blind deconvolution of adaptive optics images plays an important role in the area of space object surveillance and astronomical imaging. This paper proposes a high resolution blind restoration method for adaptive optics images based on wavelength diversity. Firstly, we utilize Zernike polynomial to establish the relation between point extension functions of the two channel images under the frame of Fourier optics theory. Secondly, we derived the general cost function of the difference between the object image and the collected images by using of the least square principle. The point spread functions of each image could be estimated by minimizing the cost function, and the object image could be obtained subsequently. The algorithm proposed in this paper can fundamentally improve the ill-posed problem of blind deconvolution. Experimental results show that our method is robust and can effectively improve the imaging quality of turbulence degraded images.

Camera calibration is one of the key steps of 3D reconstruction. Aiming at improving the accuracy and solving the complex problems of camera calibration, based on the camera model, a practical approach of camera calibration based on an isosceles right triangle planar pattern was presented. The camera coordinate system was set as the world coordinate system. Five camera intrinsic parameters needed to be calibrated. The planar pattern contained three apexes of the triangle and two arbitrary points on two right angle sides. More than six pattern images were captured from different poses, the improved Harris detector was used to obtain the sub-pixel coordinates of the corners. Established the constraint equations according to the right angle, camera intrinsic parameters can be calculated. This method reduced the requirements of pattern and was simple and convenient to operate. Experiment shows that the approach is accurate.

Fixed pattern noise (FPN) in infrared images seriously degrades the imaging quality and visual effect of infrared focal plane arrays (FPAs). Although many scene-based non-uniformity correction (NUC) algorithms have been developed recent years, the convergence speed of the bias and gain correction parameters still need to be further improved. In this paper, we present a novel NUC approach for IR FPAs which minimizes the total variation of the estimated IR irradiance guided by a noise model image, and we name it guided total variation (GTV) NUC method. A temporal detection factor is introduced to NUC procedure to prevent NU parameters updating when scene movement stops. In the proposed scheme, the correction parameters of the FPN are estimated via an iterative optimization strategy, frame by frame. The experimental results of synthetic and real IR videos demonstrate that the proposed algorithm have better NUC performance in terms of fewer ghosting artifacts and faster convergence than the state-of-the-art methods.

With the increasing demand in visual experience, visual effects with good immersion and stereoscopic sense have become one of the research hotpots in computer vision and computer graphics. This paper presents a stereoscopic panoramic image generation method based on static light field capture. The work can be divided into two parts: the acquisition of the static light field and the stereoscopic panoramic image generation. The light field data is recorded by two mutually perpendicular cameras rotated around a fixed axis. The two cameras capture images with every 9 rotating degrees and a total of 40x2 images were obtained after spin a circle. Both two cameras are calibrated and the recorded images are re-projected to a spherical surface. The optical flow information was calculated from the overlapping part of the spherical projection using a block matching method based on pyramid model. As a result, the light filed information which contains texture and disparity information is acquired based on the spherical expression. On the basis of the above work, the stereoscopic panoramic images under the new viewpoints were then resampled densely based on a pixel level from the obtained light field. A Poisson blending algorithm is used afterward to fuse the image strips from different perspectives and eliminate the brightness difference between those different strips. The high quality stereoscopic panoramic images with an 8k resolution were achieved in the end. Compared with the commonly used panoramic image generating method, our method has the advantages of low cost, simple operation, and high automation. More importantly, it can generate stereoscopic panoramic images containing parallax information with the vertical 180 degrees of view.

In this paper, we propose a new automatic laser sheet calibration in laser triangulation. The existing methods are relatively monotonous, inefficient, and subject to a large number of human interventions. Our new method has three advantages. First of all, the new method is an automated method while the traditional one is manual; Secondly, the new method simplifies the calibration process and costs fewer time; Finally, the new method can obtain higher accuracy result. We calibrated the parameters of camera and laser plane by the new method, and calculated the height of corresponding points on laser line by laser triangulation. The validity of our new method is verified by analyzing and comparing the error of height value

Mushrooms, as a delicacy in people's lives, are deeply loved by people, and the nutrients in mushrooms play an essential role in people's health. However, the characteristics of poisonous mushrooms and non-toxic mushrooms are extremely similar, and they are easily confused in the field of miscellaneous circumstances, and therefore often cause the eaters to ingest poisoning. The identification of poisonous mushrooms is a basic measure to avoid poisoning. At present, the methods for identifying poisonous mushrooms mainly include shape recognition method based on folk experience, chemical analysis methods, and animal testing methods. However, these methods have some disadvantages such as low accuracy in the practical application identification, complex experimental equipment required, unsatisfactory detection of unknown toxins, and long experimental period. Aim at the deficiency of the traditional poisonous mushroom identification method; this paper proposes a poison mushroom identification method based on BP neural network. Through the learning of the characteristics of the known poisonous mushroom, identify unknown poisonous mushrooms.

This paper presents a new strategy of handwritten digit recognition based on the improved VGG16 model that aims at the lack of texture features, the large differences between handwritings, and the difficulty of extracting effective information. In order to maximize the efficiency of the improved VGG16 model, an automatic drop-based learning rate scheduling is proposed to improve the SGD algorithm in the learning procedure. And automatic adaptation technique of the SGD optimizer's learning rate parameter according to the changes of accuracy rate in the previous training iterations is utilized for the training of the model, which not only apparently improves the learning results, but also speeds up the training convergence. The improved VGG16 model was evaluated on the extended MNIST dataset, achieving a high recognition accuracy rate of 99.97%. Experimental results demonstrate that the improved VGG16 model has obviously higher recognition accuracy than traditional classifiers, has stronger feature extraction ability and can meet the requirements of handwritten digits classification and recognition.

In order to avoid the influence of external factors on the subsequent recognition of RGB video and improve the accuracy of human motion recognition, an algorithm of human action recognition based on Two-Stream Ind Recurrent Neural Network is proposed. In terms of extracting features, the temporal network extracts the information on the 3D coordinate of different joints at each time and classifies it by a softmax layer. The spatial network converts the spatial positional relationship of the joints at each moment into a skeleton sequence and inputs it into the softmax layer to classify. Finally, the results of the classification of the temporal network and the spatial network are weighted and summed to obtain the final classification result. Experiments verify the validity of the model on the largest 3D skeleton action recognition dataset NTU RGB + D and SBU interactive dataset.

The vertical structure (VS) in nanosheet is important to electrical application because the VS in nanosheet determines the efficiency of solar cells, electrochemical biosensors and etc. The scanning electron microscopy (SEM) provides a way to observe the nanosheet structure; however, the identification of the structure is inaccuracy and when just from human justification. Deep learning gives an efficient method to identification and segmentation in an SEM image, which will enhance the precision. And the identification of VS is beneficial to get the relationship between the VS and electrical application from statistical viewpoint. In this paper we design a deep learning framework to detect the VS in the ZnO nanosheet, which overcomes two issues in the SEM: (i) the intensity inhomogeneity issue; (ii) the interference issue which is caused by other nano structures. And the experimental results exhibit height performance.

Protecting panels in substations ensure the stable operation of electrical equipment, which is of great significant to the local electrical power system. In order to reduce the manual intervention and improve the efficiency, we propose a vision-based status recognition method for protecting panels which can be used in automatic inspection equipment. The approach is divided into three stages: pre-processing, switch localization and status recognition. During the preprocessing, the image is firstly wrapped into the front view with inverse perspective transformation (IPM), with the help of a set of four artificial auxiliary marks. Meanwhile, a ROI region is extracted from the wrapped image which discards most of the trivial context. Secondly, the gradient intensity feature is calculated to locate switches on a panel. After projecting the gradient intensity image horizontally and vertically, the distribution of the switches is determined by respectively analyzing the both directional projection curves. Finally, a SVM classifier is trained to recognize the status of the switches on a protecting panel. The input of the classifier is the gradient orientation feature extracted from a normalized single switch region, and the output is the connected or disconnected states of the switch. Experiments show that our approach has low time consumption and achieves an accurate recognition rate of 99%.

Road recognition from optical remote sensing images is important for many applications like intelligent transportation system. Currently, convolutional neural networks (CNNs) methods are widely utilized in road recognition. However, many CNN methods hardly get good recognition performance when they process high-resolution images with large road width variance and complex background. For this problem, we develop a road recognition method based on a multi-scale convolutional network (MCN) with multi-level feature fusion. MCN comprises several CNNs with different scales of inputs and thus can extract multi-scale features. Each CNN fuses low-level geometrical features, middle-level features and high-level semantic features respectively from shallow, middle and deep layers. The multi-scale scheme and multi-level feature fusion make the MCN capable to handle large road width variance and complex background. Our method is validated on a manually labeled visible remote sensing image dataset. Moreover, our method is compared with CNNs without multi-scale or multi-level feature fusion and a fully convolutional network (FCN). The experimental results show that our method can well deal with complex visible remote sensing images with large road width variance.

This paper proposes a convolutional neural network (CNN)-based tree species identification method from bark image. The proposed method uses the well-known CNN model. The difficulty of our problem is to use a special tree image in which a colorful tag is stick on the bark. The tag is irrelevant to the species. In order to recognize with CNN, it is necessary to extract a region (ROI) excluding the tag. Thus, this paper proposes a ROI extraction method. Extracted ROI is fed to CNN. We evaluated the proposed method with six tree species. We carried out the evaluation experiment with various conditions, and found an optimal condition for our problem.

A method is proposed for recognition of uchen script Tibetan historical document using baseline information. The baseline is a significant feature of uchen script Tibetan characters, and the result of Tibetan document recognition can be improved effectively by using baseline information. Firstly, the baseline is detected accurately by the upper edge of character strokes in the document. Secondly, the baseline is used for segmentation of text line and character, and the text line adhesion can be handled also. Finally, the character part above baseline and that below baseline are recognized individually, and the final results are synthesized. The experiment shows that the proposed method is effective for the recognition of Tibetan historical document, and the recognition accuracy is better than that of the whole character recognition.

In recent years, deep learning has received an excellent performance in the tasks of image feature extraction and image classification. Besides, the coding-based methods have been widely focused on because of their outstanding local description. In this paper, we propose a novel method for finger-vein recognition, which combines local coding and convolution neural network (LC-CNN). Based on local graph structure (LGS), a weighted symmetrical LGS is firstly proposed to locally represent the gradient relationship among the surrounding pixels. Then, the traditional local coding methods are reconstructed with a set of fixed sparse predefined binary convolution filters. To address the over-fitting of the network, we use the local coding convolution to alter standard convolution in pre-trained CNN. Finally, the extracted feature vector are input into a support vector machine (SVM) for images classification. Experimental results show that the proposed approach achieves better performance than the traditional coding methods on finger vein recognition.

Difference summation is utilized to calculate the differential excitation of Weber Local Descriptor (WLD), however, this method is sensitive to noise and may offset positive and negative difference; at the same time, the orientation operator of WLD solely calculates the ratio of the gray difference between the vertical and horizontal orientations, that cannot portray the orientation features of the finger vein effectively. According to the features of finger vein images, this paper based on the differential excitation and orientation operator of Weber descriptors, several improvements are as below: 1)firstly, find the edge area in the image, and optimize the gradient magnitude according to the position of the pixel to increase the discrimination, 2)and then differential excitation is represented by the ratio of the optimized gradient magnitude to the current pixel; 3)double Gabor orientations are leveraged to replace the original gradient orientation to reduce the influence of translation and rotation on recognition; 4)finally, in order to better measure the similarity among features, a cross-matching algorithm is used to improve the recognition rate. In this work, the proposed local descriptor and the cross-matching algorithm are combined for finger vein image recognition, the recognition rates reached 100% and 99. 458% on the two finger vein databases at home and abroad, respectively.

Due to the underdeveloped scanning technology, some old movie films are scanned in digital format with lower resolution, which does not meet the viewing needs of contemporary viewers. Therefore, it is necessary to superresolution processing them to improve the image quality. However, some old movies will appear blurred after scanning. In this case, the existing algorithm super-resolution reconstruction results are often not ideal. This paper adds image deblurring pre-processing before the super-resolution processing. First, the old movie is deblurred according to the deblurring generation training model against the network, and then the image is super-resolution processed by the sub-pixel convolution network. The method aims to improve the problem that the repair effect caused by the image blur caused by the old film in the super-resolution reconstruction is not ideal.

Multi-sensor remote sensing image super-resolution aims to provide better characteristics for different types of resolution and compensate the limitations of the particular imaging systems. However, existing super-resolution techniques consider spectral and spatial resolution enhancement separately, i.e. only spatial or only spectral resolution can be enhanced. Among spatial super-resolution methods maximum a posteriori estimation approach with B-TV regularization stands out as one of the best method for spatial resolution enhancement. But existing implementations were designed only for RGB and grayscale photographic imagery. Unlike photographic RGB imagery, multispectral remote sensing images captured by optical sensors often contain more than three spectral channels (red, green and blue) and, moreover, different remote sensing systems produce a different spectral response for the similar spectral components. Therefore, a more complex image acquisition model should be regarded to take into account the variations in bandwidth and number of spectral channels in the case of remote sensing images. In this article, we propose an algorithm aiming to provide spectral-spatial multi-sensor remote sensing image super-resolution. We apply a joint spectral-spatial image acquisition model, that is typical for remote sensing systems, and investigate the super-resolution algorithm streaming from this model and the maximum a posteriori estimation approach with B-TV regularization. We propose a simple way to adapt B-TV regularization in the case of multiple spectral channels. Our experimental results confirm the enhancement in the spectral and spatial resolution of the output image in comparison with the input images. The results of our research demonstrate that the proposed method achieves both spectral and spatial super-resolution.

It is known that actual performance of most previous face hallucination approaches will drop dramatically as a very low-resolution tiny face is provided. Inspired by the latest progress in deep unsupervised learning, this paper works on tiny faces of size 16×16 pixels and magnifies them into their 8× upsampling ones by exploiting the boundary equilibrium generative adverarial networks (BEGAN). Besides imposing a pixel-wise L2 regularization term to the generative model, it is found that our targeted auto-encoding generator with residual blocks and skip connections is a key component for BEGAN achieving state-of-the-art hallucination performance. The cropped CelebA face dataset is preliminarily used in our experiments. The results demonstrate that the proposed approach is not only of fast and stable convergence, but also robust to pose, expression, illuminance and occluded variations.

The Tiangong-2 Space Laboratory is the second space laboratory launched by China after Tiangong-1. It is equipped with a wide-band imager with wide-band, wide-field, multi-spectral features, and has application potential in regional planning of disaster prevention and disaster reduction, disaster monitoring and loss evaluation. This paper proposes a method of residential area extraction based on Tiangong-2 images. Taking full advantage of the spectral response characteristics of typical objects such as vegetation, water body and residential area in the images, on the basis of image geometry processing and image enhancement, image threshold segmentation and remote sensing index are used to realize the extraction of feature samples. Through the NAPCA transform and supervised classification of SVM, the extraction of residential areas is realized. The method can realize the rapid and semi-automatic selection of residential area, vegetation and water samples, and can obtain better extraction precision of residential areas. The experimental results show that the Tiangong-2 images with 100-meter spatial resolution can achieve the extraction accuracy of more than 70% of the residential area, and can be used for large-scale spatial distribution analysis of residential area and disaster impact area assessment.

Image super-resolution methods based on forward-feed convolutional neural networks (CNN) reconstruct the image with more details and sharper texture. However, most of these methods do not consider the influence of high level semantic feature to improve image perceptual effect. In this paper, we propose a deep CNN architecture jointing low-high level feature for image super-resolution. Our method uses 17 weight layers to predict residual between the high resolution and low resolution image. And we joint the low level and high level image features to constraint the network parameters updating. Experimental results validate that our method reconstruct the high resolution images with clear edge and less warp.

Due to limited training data for SAR target recognition tasks, different regularization techniques have been used in the deep convolution neural networks (CNN) to improve generalization ability. In this paper, the influences of three regularization techniques, including data augmentation, L₂ regularization term, dropout, are studied under standard operating conditions (SOC) when moving and stationary target recognition (MSTAR) dataset is used for SAR target recognition. Four representative CNN models based on classical models, such as AlexNet and ResNet, are selected and trained to recognize 10-classes targets. Additionally, a CNN model which has fewer network parameters is designed based on multi-scale spatial feature extraction strategy and SqueezeNet to study the influence of the amount of network parameters. The experimental results demonstrate that, when using the AlexNet series model for SAR target recognition, using dropout may greatly improve the ability of model optimization. ResNet series models which have more layers, have better effect on Test 1+noise than other CNN models, especially taking dropout in the model. For the models based on highway networks, adding L₂ regularization terms in loss function can improve the test accuracy, but it also makes the latter phase of training extremely unstable. Data augmentation is an effective regularization technique when the model can get high training accuracy.

Land cover detection and classification has been an important component of Geographic Information Systems. They are used in policy planning, socio-economic analysis, cartography and Government scheme planning and evaluation. Our study uses high-resolution time-series satellite images of Indian cities between years 2000- 2017 and measures the changes in area occupied by infrastructure such as buildings and hutments during that period. To detect buildings and hutments we train a U-Net model¹ for image segmentation task and highlight the boundaries for man-made constructions i.e. buildings and hutments for each block in our New Delhi data collection. We have also provided sample contrast against the development information available on BHUVAN portal, made publicly available by Indian Space Research Organization (ISRO) study. Using the time-series data of building and hutment growth, we can enable urban planners and policy makers to identify necessity of supplementary resources like government hospitals, roads, gardens, etc.

In this paper, we propose a new disparity-based obstacle detection method for intelligent vehicle. The disparity maps captured by the stereo vision system ZED is utilized to detect obstacles in outdoor environment. This method consists of several steps namely, histogram calculation based on the disparity map, peaks finding using Non-Maximum Suppression (NMS), disparity slicing and obstacle detection using contours. The first step is to obtain the distribution of the pixels in the disparity map. The second step is performed to get the peaks in the histogram after the binary processing on the disparity map. However, there are always many noise objects in the binary image. The real obstacle area needs to be separated from the noise obstacles. Thus, the third step is performed to filter the binary image based on a statistics by column and then segment the detected obstacle area. The last step is to calculate the obstacle including its size, distance and position on the disparity map. A series of experiments prove that the proposed method is accurate at detecting obstacles and calculating the required information based on stereovision techniques.

For the road mark in the presence of abrasion, adhesion, and occlusion, it is difficult to accurately locate the road mark by typical point cloud clustering and segmentation method. An automatic road mark detection method based on local point cloud projection and 2D deep learning object detection is proposed. Firstly, the local ground point cloud regions are obtained by prior information; Secondly, we perform orthogonal projection on the point cloud to project the point cloud into two-dimensional image by affine transformation and use R-FCN detection method to detect the road mark in the image. The holes on the image are filtered by local maximum filtering. A line-by-line search strategy based on maximum reflection rate is proposed to refine the detected bounding boxes to improve the detection accuracy; Finally, we implement inverse affine transformation and local coordinate search to restore the road mark coordinates from twodimensional image to 3D point cloud. In the point cloud dataset collected from China highway, our experimental results show that the proposed method can achieve road mark detection under complex scenes such as occlusion, abrasion and adhesion. Comparing with other methods, the detection recall, accuracy and efficiency have been improved greatly.

In video foreground detection, the frame difference method bears a fast detection speed and strong timeliness. However, the detection foreground target is not complete enough, tending to result in voids and poor robustness. Although the Gaussian mixed model does well in detection, a ghost image is easily brought forth when it starts foreground target detection towards sudden motion. As far as these problems are concerned, a foreground detection algorithm based on improved Gaussian mixed model is proposed in this paper. The foreground region detected by the Gaussian mixed model is matched with the one detected by the improved three-frame difference method with the matched foreground reserved. Then the unmatched one is regarded as a 'ghost' region. The background model of the region is updated and the mean of the maximum weighted Gaussian model is replaced in an usage of the pixel of the corresponding area, thus breaking the obstacles in traditional method for detecting holes and 'ghosting' problems. The experimental results have shown that the proposed algorithm has better robustness and accuracy in different backgrounds, and the precision and recall excel that of traditional algorithms.

Object detection is one of the most popular and difficult field in computer vision. Although deep learning methods have great performance on object detection. For specific application, algorithms which use hand-crafted features are still widely used. One main problem in object detection is the scale problem. Algorithms usually use image pyramid to cover as many scales as possible. But gaps still exist between scale levels in image pyramid. Our work extends some sub scale level to fill the gaps between image pyramids. To this end, we use Gaussian Scales Pyramid to generate sub-scale image and extract HOG feature on the sub-scale. We use framework offered by DPM algorithm and make modification on it. We compare the result of our method with DPM baseline on Pascal VOC database. Our work has great performance on some categories and makes an improvement on the overall performance. This work can be used in other object detection frameworks. We apply multi-scale HOG feature on pre-process procedure of our own detection framework and test it on our own dataset. Then the framework gains performance improvement on precision and recall rate of the pre-process procedure comparing to the original HOG feature architecture.

Object detection is the basic research direction in the field of computer vision. It provides basic image information data for other advanced computer vision processing and analysis tasks. With the continuous breakthrough of deep machine learning technology, especially convolutional neural network model in the field of digital image processing shows a strong ability to extract image features. By choosing the depth separable convolution layer to replace the standard convolution layer used in the traditional model, the number of parameters of CNN network model is compressed. Depth Separable Convolution Layer (DSCL) decomposes the standard convolution layer factor into depth convolution layer and point convolution layer, and extracts and merges image features in two steps to reduce the number of parameters. By introducing a depth-separable convolution layer instead of a standard convolution layer, the number of parameters of the model convolution layer is reduced by 78.1%. We choose image feature pyramid network to fuse the image features extracted from each layer of CNN network, so that the target detection model can use matching image fusion features for different size and shape of the target to be detected. The average detection precision on the PASCAL VOC dataset increased to 77.5%.

Yarn-Dyed fabric defect detection is an important part of the textile production process, in which rapid and accurate detection is the main challenge in textile industry. However, the performance of defect detection largely depends on whether the manually designed features can properly represent the features of the defects. In this paper, a new detection algorithm for automatic fabric defect detection using the deep convolutional neural network (CNN) is put forward. Our defect detection algorithm is based on three main steps. In the first step, a preprocessing stage decomposes the fabric image into local patches and labels each local patch accordingly. In the second step, labeled fabric samples are transmitted to deep CNN for pre-training. Finally, defects are detected during image inspection that trained classifier slides over the entire fabric image and returns the category and position of each local patches to achieve defect detection. The proposed method was validated on two public and one self-made fabric databases. By comparing manually designed image processing solutions with other deep CNN networks for feature extraction methods, the experiments show that the proposed method can inspect defects at a higher accuracy compared with some existing methods.

The main goal of object detection is to recognize and locate the object of interest from the static image or video sequence. It is one of the key tasks in the field of computer vision. However, there are many factors in brightness, shape, color and occlusion of targets, and they are disturbed by complex environmental factors, which make the research opportunities and challenges of object detection algorithms coexist. In this paper, two main frameworks of object detection algorithm based on convolutional neural network are researched, which are based on region proposals and regression idea respectively. Then we present a joint mechanism algorithm for object detection. This algorithm makes a balance between detection efficiency and accuracy to make it more meet the actual needs. The internal of the algorithm is adjusted and optimized, so that the two detectors can make their own judgments according to the characteristics of the image, and decide whether to detect the object to classify and locate it, so that the efficiency is higher and the accuracy is also improved.

Pedestrian detection (PD) is an important application domain in computer vision and pattern recognition. Conventional PD in real life scene is usually based on a fixed camera, which can detect and track the pedestrians in the monitoring region. However, when the pedestrian leaves the visible area of the fixed camera, it is usually difficult, if not impossible, to monitor the pedestrian. In response to the limitations of the conventional pedestrian detection application scenarios, a four-rotor unmanned aerial vehicle (UAV) system equipped with a high-definition (HD) camera is designed and implemented to detect human targets. Considering the size of human body in aerial image is small and easily to be occluded, we draw on the advanced research results in the field of target detection and propose a robust pedestrian detection method based on YOLO (You Only Look Once) network. The flow of the proposed approach is as follows. Firstly, the HD camera, which is installed on the monitoring UAV, is used for capturing images of the designated outdoor area. Secondly, image sequences are collected and processed using the airborne embedded NVIDIA Jason TX1 and Ubuntu as the core and operating system, respectively. Finally, YOLO is used to train the pedestrian classifier and perform the pedestrian detection. Experimental results show that our method has good detection results under the complicated conditions of detecting small-scale pedestrians and pedestrian occlusion.

Aiming at the problem of small moving object detection on dynamic water surface background, a moving object detection algorithm of interval background model is proposed. The algorithm first learns the range of background values on each coordinate point in the video sequence from a video of a non-moving object, and forms an interval background model. when the moving target is detected, the detection is carried out by point-by-point scanning, the value of the current pixel is compared with the value of the background model, if the value of the pixel point is within the range of the background module, the pixel point is the background, otherwise, the moving object is the background. After all pixel points are detected, they are combined into an initial moving target image. Finally, initial movement The target image is subjected to de-interference processing to obtain a final moving target image. The experimental results show that the algorithm has good detection effect on dynamic water surface moving targets and is fast in speed.