To meet the real-time requirements of fabric defect detection in industry, a real-time fabric defect detection algorithm based on S-YOLOV3 (Slimming You Only Look Once Version 3) model is proposed. To develop this algorithm, the K-means clustering algorithm is used to determine the target prior frame for adapting to different sizes of defects. The YOLOV3 model is then pretrained to obtain the weight parameters, and the scaling factor γ is used in the batch normalization layer to evaluate the weight of each convolution kernel. The convolution kernel with weight value lower than the threshold is clipped to obtain the S-YOLOV3 model to achieve compression and acceleration. Finally, the pruned network is fine-tuned to improve the model detection accuracy. Experiment results reveal that the proposed model provides accurate detection of fabrics with different complex textures (average precision of 94%). After pruning, the detection speed is increased to 55 FPS. The obtained accuracy and real-time can meet the actual demand of industry.

To address the issue that large amount of data processing and poor performance of super-resolution reconstruction of license plate images in surveillance video, in this paper, a super-resolution construction algorithm of license plate image based on gradual back-projection network is proposed. First, in order to reduce the amount of super-resolution network data processing, the low-resolution license plate area is detected and extracted. Second, the larger sampling multiples of deep back-projection network (DBPN) are decomposed and the iterative back projection is completed by sampling step by step. In the gradual back-projection unit, skip connection fuses the middle-scale features produced by the gradual sampling operations to improve the feature utilization rate. 1×1 convolutional layer is used for reducing dimensions of fused middle-scale features, while preserving key information. Finally, the high-resolution license plate image is reconstructed according to the feature image generated by the gradual projection unit. Experimental results show that the proposed algorithm not only reduces the amount of data processing and parameters of the super-resolution network, but also greatly improves the subjective feeling and objective evaluation index of the reconstructed license plate image quality compared with DBPN.

In order to solve the problem of color offset distortion and halo effect when using traditional dehazing algorithms to process the images with large areas of sky and abrupt changes of scene depth, an image dehazing optimization algorithm based on transmittance adaptive constraint correction is proposed. The algorithm combines automatic and manual estimation in the atmospheric light value estimation stage, which is convenient for users to further adjust the dehazing results according to their needs. For the estimation of transmittance, first, the lower limit of the estimation of transmittance is obtained through the boundary constraint of the scene radiance to replace the fixed value set in the traditional algorithm. Then, the threshold value is set to determine whether the pixel is within the same depth of scene, and the corresponding adaptive correction is made according to the intensity difference ratio to optimize the estimation of transmittance. The results show that this algorithm can achieve a better image dehazing effect for the haze image. It can not only restore the clear image, enhance the visual effect, and usability of the images, but also effectively avoid the color deviation artifact in the bright area of the image and halo effect in the sudden changes of depth of areas.

The traditional fire smoke detection method has a degraded detection performance in the case of complex scenes and high interference. Aiming at this problem, this paper proposes a convolutional neural network fire smoke detection method based on target region. A two-layer fire smoke detection model is constructed. Using the motion detection algorithm of the target region positioning layer, the smoke target region is extracted from the fire smoke image, which can quickly remove a large amount of irrelevant interference information in complex scenes, and input the extracted smoke target region into the fire smoke recognition layer, and then extract the deep features of the smoke through the convolutional neural network to classify it to complete the fire smoke detection. Experimental results show that the proposed method has strong anti-interference performance in the data set under complex scenes, which effectively reduces the false detection rate and improves the accuracy of smoke detection.

Deep learning has impacted the research and application of face recognition to some extent; however, it is unsuitable for small embedded devices owing to its large computational cost and time consumption. Herein, a facial feature extraction method for integrating gradient features in a lightweight convolutional neural network (SqueezeNet) was proposed to ensure the application of the network model to embedded devices with relatively small memory and facial features that are more robust to different lightings. Experimental results showed that the lightweight convolutional neural network integrating the first-step gradient feature extracted by dividing the image into a block of 8 × 8 can achieve a recognition rate of up to 97.28% in LFW dataset, which is 4.36% higher than that of the conventional lightweight convolutional neural network.

To address the problem of difficulty in eliminating the ghost in conventional ViBe algorithm in moving object detection, an algorithm based on Euclidean distance and Tanimoto coefficient is proposed. The proposed algorithm calculates the similarity of gray histogram of foreground and neighborhood pixels to detect and eliminate the ghost. Moreover, the conventional ViBe algorithm cannot eliminate the shadow in moving object detection. To mitigate this issue, an algorithm combining YCbCr color space and Gaussian mixture shadow model is proposed to detect and eliminate the shadow. Simulation results reveal that the proposed algorithm can effectively eliminate ghosts and shadows while maintaining the efficiency of the conventional ViBe algorithm.

To reduce the interpolation error in parallel phase-shifting digital holographic reconstruction and improve the quality of the reconstructed image, an interpolation algorithm of parallel phase-shifting digital holographic fringe analysis is proposed on the basis of holographic image fringe features and pixel distribution. First, holographic images with the same phase are extracted from the parallel phase-shifting digital hologram according to the phase-shifting characteristics, and each image is divided into 3×3 square areas. Then, multiple-direction interpolation is carried out in the divided square areas according to the holographic fringe characteristics. The image frequency-domain characteristics after multiple-direction interpolation are combined selectively, and the newly generated holographic field is reconstructed from the merged image. Experimental results show that the peak signal to noise ratio of the reconstructed image obtained by the proposal method is more than 45% higher than that of the traditional parallel phase-shifting holographic interpolation algorithm. And the computing time can be reduced to about half of that of other interpolation algorithms. The proposed algorithm can better reconstruct parallel phase-shifting digital holographic images, reduce the interpolation error, keep the detailed information, and improve the quality of the reconstructed image. The proposal algorithm can be applied to obtain three-dimensional field information in many fields, such as biological cell observation, physical examination for moving object, and microscopic particle imaging and tracking.

To improve the detail information and preserve the contrast of fusion images of visible light images and infrared images, a multi-scale decomposition image fusion method based on residual learning and visual saliency mapping is proposed. First, a Gaussian filter and a guided filter are used to perform multi-scale decomposition, therefore decomposing the image into a basic layer and a detail layer. The detail layer is divided into a small-scale texture layer and a middle-scale edge layer. Then, the proposed improved visual saliency mapping method is used to fuse the base layers, and the base layer of the low-light image is enhanced to make the fused image have good contrast and overall appearance. For the detail layer, a residual network deep learning fusion model is proposed to maximize the fusion rules of the small-scale texture layer and maximize the middle-scale edge layer, respectively. The experiment compares the proposed algorithm with the latest six methods on the four objective indicators of discrete cosine feature mutual information, wavelet feature mutual information, structural similarity, and artifact noise rate on the TNO dataset. The proposed algorithm is improved in the first three objective indicators and further decreased in the artifact noise rate. This algorithm not only keeps the salient features of the image, but also makes the fused image have more detailed texture information and good contrast, therefore effectively reducing artifacts and noise.

Nonrigid image registration technology is an important research direction in the computer field. It is widely used in medicine, astronomy, and military applications. In the process of nonrigid image registration, it is often impossible to simultaneously deal with local large and small deformations. To solve this problem, in combination with the theory of normalized mutual information image registration, an Active Demons nonrigid image registration algorithm based on mutual information is proposed herein. This algorithm improves the driving force of the Active Demons algorithm by introducing mutual information, which is used to register the image. The original image is used to adaptively adjust the normalization factor, and then the algorithm conducts test simulations on medical, natural, and synthetic images. Experiment results show that the proposed algorithm is superior to the original Active Demons algorithm in terms of registration accuracy and robustness. It can process large deformed images and retain their edge features, thereby providing better registration accuracy and visual effects.

To address the problems of color distortion, low contrast, and blurred vision in degraded underwater images, an underwater image enhancement algorithm based on high and low frequency component fusion is proposed. First, multi-scale retinex algorithm is used to estimate high frequency components. Then, contrast-constrained adaptive histogram stretching is performed to enhance the global contrast while highlighting details. To prevent noise generated during image stretching from affecting the image quality, guided high frequency components are denoised via guided filtering. Then, the original image and high frequency components are divided to obtain low frequency components, and the multi-scale detail extraction method is used to obtain detailed information. Finally, the noise-removed contrast-enhanced image and the high and low frequency detail image are linearly weighted and color corrected to obtain the final underwater clear image. Experimental results show that the proposed algorithm can effectively enhance the image contrast and details and significantly improve the visual effect of the image.

Focusing on the issue of ambient light interference affecting a depth camera and difficulty in obtaining the tree depth images, tree depth image generation algorithm based on small motion short video images is proposed. The algorithm tracks and matches the subpixel corners obtained after segmentation, uses the bundle adjustment to estimate the camera parameters. Intensive stereo matching and denoising of image sequences using plane scanning method to obtain tree depth images. The effectiveness of the proposed algorithm is verified using smartphones by collecting short video images of 1-2 s small movements, and the generated tree depth images exhibited a significant improvement in subjective effect. Verification on the depth image database NYU depth v2 revealed root mean square error is 60.58 and relative error is 0.34. Experiment results show that using this algorithm in a natural setting, a fine tree depth images can be effectively generated without the need for depth cameras, camera calibration, and training of a large number of RGB (Red, Green, Blue) images and depth images, which can reduce the data collection and storage costs. The research findings can provide a reference for the visual reconstruction of standing wood and measurement of the standing wood factor.

Mural image has the characteristics of rich structural details, complex textures and variable colors, while the mural image reconstructed by image super-resolution algorithm based on convolution neural network has problems of texture blur and edge staircase effect. Therefore, we propose a super-resolution reconstruction algorithm based on multi-scale residual attention network. First, the features of low-resolution mural images with convolution kernels of different scales are extracted directly by multi-scale mapping unit. Then, the fused feature maps are input into the attention block of residual channel , so that the weight of each feature map is optimized from the global information, and the depth mapping ability of the network model is enhanced. Finally, a sub-pixel convolutional layer is introduced at the end of the network to rearrange the pixels to obtain the reconstructed high-resolution mural image. Experimental results show that this algorithm can reduce the reconstruction error, enhance the edge and structure information of the reconstructed mural image, and enrich the texture details of the reconstructed mural image.

When extracting foreground targets in the complex background with dynamic interference factors, the existing algorithms of extracting foreground targets in visual background are prone to ghost image and false detection, so an improved algorithm based on visual background is proposed in this paper. First, according to the time series and position characteristics of pixels, the matching probability, matching degree, and brightness information of the pixels are calculated. Second, background model matching the current complex background is updated in real time, and the background model is initialized. Finally, the video in various complex backgrounds in the CDnet 2014 dataset is tested, and compared with the classic Gaussian mixture model, visual background extraction (ViBe) algorithm, and improved ViBe algorithm. Experimental results show that the algorithm can efficiently remove the effects of ghosts in various complex backgrounds, had a high extraction precision, which greatly reduces the misclassification rate and missed detection rate of the extraction results, and improves the efficiency and robustness of the algorithm in complex background.

Adjacent picture elements division of focal plane (DoFP) polarizing camera have different polarization directions, which will result in a loss of imaging resolution. Bilinear interpolation and sliding window algorithms can improve the imaging resolution, therefore, this paper analyzes the mathematical relationship between bilinear interpolation and sliding window algorithms, and degree of linear polarization (DoLP) values are calculates after processing at the edges. The analysis shows that the Stoke vector obtained by the bilinear interpolation algorithm is the average value of the processing results of by sliding window algorithm in four directions. In the edge region, the DoLP processed by sliding window algorithm is larger than bilinear interpolation algorithm. We apply the two algorithms are respectively applied to raw DoFP polarization images and evaluated by visual, statistical histogram, and mean value of the DoLP image. Experimental results prove that compared with bilinear interpolation algorithm, the sliding window algorithm produces stronger pseudo-DoLP information at edges, which highlights the details of the edges.

To address the limitations of existing detection methods in pulmonary nodule detection, such as low accuracy and over-fitting, a pulmonary nodule detection method based on an improved YOLACT model was proposed. In the main structure of the YOLACT model, the original residual network was replaced with DetNet to overcome the limitation of the original model in small nodule detection. Further, a transfer-learning mechanism was introduced in the model training to prevent the over-fitting problem of the original model induced by learning difficulties on a small number of pulmonary nodules, thereby allowing the new model to achieve better detection results. Moreover, the original ReLU function was replaced with the RReLU function to further reduce the possibility of over-fitting. Experimental results on LUNA16 dataset indicate that the proposed method can achieve improvement under several evaluation metrics, such as the working curve of the subject, rate of false positives, rate of missed diagnosis, and accuracy.

This study proposes a multi-scale key point extraction algorithm based on normal weighting to address the sensitivity to noise and dependency on object models' shape features in traditional keypoint detection algorithms. First, at each scale, the covariance matrix of the local neighborhood is established and the ratio of the local coordinate system appearing on the first two axes is calculated. Thus, candidate keypoints are determined based on the ratio. Then, to measure the local maximum dissimilarity measured value of the point cloud, the normal weighted shape index value is calculated. Finally, the maximum value point of the local maximum dissimilarity measured value at different scales is selected as the final keypoint. The experimental results show that compared with other traditional algorithms, the proposed algorithm can effectively extract keypoints of various point cloud models and simultaneously consider the quality and quantity of keypoints and operating efficiency. Moreover, the proposed algorithm has strong adaptability for models with sharp features and large area smooth features, which enhances its robustness and shape index function.

This paper proposes a sorting method for non-ferrous metal fragments based on machine vision to overcome the problems of complicated source, difficult sorting, and low recognition accuracy rate of waste non-ferrous metal fragments. Using color moments and Tamura texture characteristics, the optimized support vector machine (SVM) sorting algorithm based on principal component analysis (PCA) is established. Thus, a new concept of high-precision automatic sorting from the machine vision perspective is proposed herein. Results show that the proposed SVM algorithm based on color and texture features can effectively detect and classify metal fragments with an accuracy of 93.89% and improve the recognition speed. The proposed algorithm meets the requirements of large-scale and efficient separation of scrap metals.

Aimat the problems of Poisson noise and uneven illumination in the visible light image of pantograph-catenary arcing, we propose an image processing algorithm including noise reduction and region segmentation. First, Poisson noise in pantograph-catenary arcing image is removed by two-dimensional Gabor wavelet transform algorithm, and then the illumination uniformity is judged according to the improved uniformity measurement algorithm. For the image with uneven illumination, two-dimensional Gamma function is used to segment. Simulation results show that the algorithm can effectively suppress Poisson noise, accurately segment the target area, and the segmentation accuracy reaches 95%. It provides an idea for quantitative analysis of intensity of pantograph-catenary arcing.

Aiming at the problems of information loss, unclear region and fog occlusion in foggy images, in this paper, a Gaussian pyramid transform Retinex image enhancement algorithm based on bilateral filtering is proposed to improved the contrast of foggy image. First, an improved mathematical model of bilateral filtering function based on spatial kernel function and pixel difference is established. Second, the input image is converted into HSI (hue, Saturation, intensity) image, and the improved bilateral filtering function is used to replace the Gaussian function in the original algorithm. The reflection component is extracted from the luminance image (I color space) to obtain a reflection image with edges retained and unaffected by brightness. Finally, color images of different scales are obtained through Gaussian pyramid down sampling, the images of different scales are enhanced by improved Retinex algorithm, and the image is reconstructed based on Gaussian Laplacian algorithm to improve the image contrast. Experimental results show that the algorithm can effectively enhance the contrast of the image, and the color of the processed image conforms to the observation effect of human eyes.

Aiming at the problems that the classical bilateral filtering algorithm has a poor effect on depth image repair and cannot adjust the kernel function parameters accurately, a depth image repair algorithm based on morphology and improved bilateral filtering is proposed. First, the morphological algorithm is used to optimize the holes in the depth image to fill some small holes and filter out random noise. Then, using the improved bilateral filtering algorithm, the probability distribution function and the maximum likelihood function are introduced to calculate the kernel function parameter values in the neighborhood of each cavity and thus to adjust the kernel function parameter adaptively and realize the repair of large area holes. Finally, the median filter algorithm is used to smooth the image and thus to remove the "burr" of the depth image, and the edge details of the image are retained and the sharpness is also maintained. The experimental results show that the proposed algorithm can effectively fill the holes in the depth image without losing the original depth image information, can realize the goal of edge preservation and denoising, and has strong robustness.

To solve the problem of uneven illumination and color distortion in the Retinex algorithm for defogging, we propose an enhancement algorithm of color images with fog based on low illumination. First, this algorithm converts the red-green-blue (RGB) image into the hue-saturation-value (HSV) spatial region, extracts the value (V) component, and performs Gamma correction for the V component after the single-scale Retinex algorithm is applied to the V component. Then, the Gaussian filter in the MSRCR algorithm is changed into guided filtering and the low-pass filtering is performed. Finally, the images obtained by the improved SSR algorithm, the MSRCR algorithm, and the Retinex algorithm based on the Laplace pyramid are weighted and fused. The proposed algorithm can get a good effect in fog removal and can effectively suppress halo and improve color distortion. As for those images processed by the proposed algorithm, the image similarity, information entropy and other indicators have been improved.

In order to further improve the classification accuracy of ship classification method for point cloud images, a new ship classification method based on three-dimensional convolutional neural network (3D CNN) is proposed. First, the point cloud image is transformed into a voxel grid image by the density grid method and the voxel grid image is taken as the input object of a 3D CNN. Then, the high-level features of the voxel grid image are extracted by the designed 6-layer 3D CNN to capture its structural information. Finally, the classification results are obtained using the Softmax function in the output layer. The experimental results show that the classification accuracy of the proposed method can reach 96.14% on the self-build point cloud image ship dataset, 5.97% higher than that of the 3D ShapeNets method and 2.46% higher than that of the VoxNet method. Compared with some existing methods, the proposed method has higher classification accuracy on Sydney urban object dataset. These results show that the proposed method has a good classification performance.

As for the problems such as the inaccuracy of transmission estimation and the color distortion of sky areas or large white area using dark channel prior dehazing, we propose a single image dehazing method based on superpixel segmentation combined with dark-bright channels. First, the superpixel method is used to segment the hazy image, and the obtained superpixel block replaces the fixed square filter window of the dark channel. Second, the prior method which combines dark and bright channels is used to obtain the atmospheric transmittance, and thus the transmittance estimation is more accurate. Thirdly, the atmospheric light value is determined by threshold segmentation combined with the bright channel prior theory in the sky region, and subsequently the transmittance is optimized by the guidance filter method with gradient information. Finally, the hazy image is restored to the dehazed image based on the atmospheric scattering model. The experimental results show that the transmittance and the atmospheric light value estimated by the proposed method are accurate, and a good dehazing effect can be obtained. The proposed method is superior to other comparison algorithms in subjective and objective evaluations.

This paper proposes a fall detection algorithm based on convolutional neural network and XGBoost. The YOLO-v3 algorithm based on the squeeze-and-excitation block is used to detect the human body area of the picture. Then, the human body pose estimation network is used to obtain the human body joints and feature vectors. Finally, we input the feature vectors into the XGBoost for training to determine whether the human body falls. The experimental results show that the proposed fall detection algorithm has a high accuracy of 98.3%.

Accurate and efficient identification of bronze inscriptions is of great significance in archaeology, history, philology and other fields. We construct the publicly available bronze inscription data set to provide data reference for subsequent research. At the same time, on the basis of the convolution kernel pruning based on L1 norm, an automatic model pruning strategy is proposed. The experimental results show that if taking the LeNet network as the basic model, the number of parameters and the FLOPS after pruning are only 30% and 69% of those for the original model, and the accuracy is up to 97.62%. The proposed method can be well adapted to the bronze inscription data set.

Images captured underwater often suffer from haze, noise, and low contrast owing to the absorption and scattering of water and suspended particles, making it difficult for analysis and understanding. To overcome these limitations, combined with an underwater optical image formation model, a fast variational approach based on a Laplace operator prior term is proposed herein to simultaneously perform dehazing and denoising. Based on the underwater optical image formation model, the data and regular items of the unified variational model are designed, wherein the Laplacian operator prior term is adopted as the regular term. The prior estimation of the improved red channel and the underwater red channel are used to obtain the global background light and the transmission map, respectively. To further accelerate the whole progress, a fast alternating direction multiplier method (ADMM) is introduced to solve the energy function. Our proposed variational method based on the Laplace operator prior term is executed on a set of representative real underwater images, demonstrating that it can successfully remove haze, suppress noise, and improve contrast and visibility.

In the field of health care, curved surface light sources are often required to illuminate the surface of the test object. It is of research significance to design a light-emitting diode (LED) array to produce uniform illumination distribution. In this paper, the cylindrical LED surface array is divided into two parts, a linear LED array and a circular LED array. The objective function distribution is constructed for the light field distribution of the curved LED array source and the column-shaped surface target receiving surface. The illumination uniformity of the illumination of the annular LED array is determined by the optical radiation parameters of the LED lamp beads, the lamp array surface radius, the receiving surface radius, and the angle between adjacent two lamp beads. The illuminance distribution of the two arrays is calculated by TracePro ray tracing method. By optimizing the arrangement of the LED surface array, a uniform illumination distribution is obtained on the cylindrical target surface.

Space cameras are the major devices used for space situational awareness and their imaging quality therefore needs to be monitored in real time. However, rotating satellite platforms to observe ground targets to measure the modulation transfer function (MTF) increase the technical risk and cannot perform multiple measurements. Thus, research is required for developing a method without platform flipping for real-time measurement of the MTF of the space camera on-orbit. The final point-based point MTF detection method utilizes the common star points in the deep sky background as the target of the on-orbit MTF measurement, combined with the point source method to study its on-orbit MTF measurement method, and utilizes Gaussian fitting positioning and multi-graph registration to reconstruct the point spread function. Simulation and experimental results show that the proposed method has desirable noise immunity. When the standard deviation of random noise is 40, the mean square error of the MTF is in the order of 10 -4, and the detection effect of small star targets is good. When the mean square error of MTF is less than 0.2 pixel, the smallest radius of the circle of confusion that can be detected is 0.9 pixel.

The detection angle within X-ray fluorescence computed tomography (XFCT) is closely related to fluorescence collecting efficiency. We design a multi-detection angle pencil-beam XFCT imaging system base on MCNP5 to simulate the imaging of a gold nano-solution columnar phantom with a mass concentration of 1%, and use filtered back projection (FBP) algorithm and joint simultaneous algebraic reconstruction technology (SART) and maximum likelihood expectation maximization (ML-EM) algorithm to reconstruct the element distribution. The contrast-to-noise ratio is calculated to quantitatively analyze reconstruction image quality with different algorithms at different angles. The results show that the FBP algorithm has better imaging quality at the backscatter angle. The SART algorithm and the ML-EM algorithm have better imaging quality at the vertical angle and the backscatter angle.

To realize the lane marker line identification in complex variable light environment and ensure all-weather lane departure warning, a new lane marker line identification algorithm is proposed. The adaptive image segmentation technology based on OTSU algorithm is used for the lane marker line division for different lighting conditions, and then the adaptive threshold is obtained the weighting process for global and partial thresholds. Sobel operators with gradient directions of 45°and 135° are adopted to extract the lane edge information. Finally, improved Hough transform method is used to finish the lane marker line identification. The road images under different lighting conditions are compared, and the results show that the identification accuracy of the proposed method is improved by 5.7% on average compared with that of the traditional Hough transform method, and the average detection time of a single image is 57.79 ms. The proposed algorithm has good anti-interference performance, can adapt to various lighting conditions, and can meet the real-time requirements of the system.

Aiming at the problem of low matching accuracy and matching failure when using point features or line features alone in remote sensing images, a fully automatic registration algorithm for remote sensing images incorporating point and line complementary features is proposed. First, the improved scale-invariant feature transform (SIFT) algorithm is used to obtain the initial matching points, and the normalized cross-correlation (NCC) measure and the random sampling consistency algorithm are used to eliminate possible mismatches to obtain the points with the same name with high accuracy. Then, an improved line segment detection operator (LSD) is used to extract line segment features, determine candidate matching line segments and construct feature descriptors by known homography geometric transformation constraints and slope constraints, and then obtain line segments with the same name. Finally, the intersection point of the line segments with the same name is extracted, and the same-named point set of the first step is integrated to calculate the projection transformation parameters between the images to realize the image registration. Experimental results show that the proposed algorithm has significant advantages in matching accuracy and matching accuracy.

The aero engine blade works in the culvert of high temperature and pressure for a long time, which leads to the damage of fracture, scratch and deformation. Three-dimensional digital modeling of damaged blade is the key step of its additive manufacturing. The accuracy of the model directly affects the quality of cladding manufacturing. In view of the spatial torsional characteristics of the space engine blade, in this paper, a non-contact laser overlapping three-dimensional reconstruction method is proposed. Based on the feature matrix of the relative position between the laser probe and the blade in the scanning strategy. The registration parameters of the point cloud are initialized, and the size of the source point cloud is reduced by assigning the weight, and the iterative nearest point algorithm is optimized. Finally, the point cloud data is simplified and meshed to complete the three-dimensional model of damaged blade. In the registration process, the effect of point cloud block overlap rate on the model accuracy are analyzed. Experimental results show that when the overlap rate is 50%, the acquisition time is short, and the information integrity is high. The standard measurement block ladder verifies that the measurement accuracy of the model is 10 μm. The method meets the accuracy requirement of the damaged blade reconstruction model.

In this study, a tracking algorithm based on channel pruning and weighting is proposed for improving the speed and accuracy of the convolutional correlation filter algorithm. This algorithm selects the single-layer convolutional features that are suitable for tracking an object. Initially, the feature mean ratio is proposed to prune the inconclusive channels; then, a combination of one-dimensional gray features is used for improving the feature representation. Subsequently, we construct the weighted correlation filter algorithm by considering the feature mean ratios as the convolution channel weight for predicting the target position. Further, an accurate location method based on the minimizing mean frame is used to reduce the prediction location error. Finally, the tracking model is updated to improve the tracking speed. The different algorithms are tested using the OTB-100 dataset. Results show that the average distance precision and the average speed of the proposed algorithm are 91.3% and 31.8 frames/s, respectively. Furthermore, the proposed algorithm can track an object under occlusion, scale variation, fast motion, and deformation in real time, effectively improving the speed and accuracy of object tracking.

Objectdetection in complex environment is affected by many factors. Traditional robust principal component analysis (RPCA) fails to obtain the lowest rank representation from disturbed data. Therefore, a novel method of video denoising and object detection algorithm based on RPCA model with l1-total variational (TV) regularization constraints is proposed. Based on RPCA, under the framework of low-rank sparse decomposition, the low-rank nature of the nuclear norm is used to model the background, and the three-dimensional TV combined with l1 norm regularization to constrain the sparsity and spatial continuity of the foreground object, and then l2 norm regularization is combined to constrain the noise part so as to make up for the deficiencies of the existing RPCA model. Using alternating iteration method, augmented Lagrange multiplier method is used to optimize the objective function, and the denoising and target detection in complex environment are realized. Experimental results show that the method can not only accurately detect moving objects under noise interference, but also maintain a relatively fast running speed, which provides a reference for the real-time detection of video. Compared with other similar methods, it not only has better detection effect, but also has advantages in the three indicators of quantitative evaluation.

To solve the problem of low classification accuracy of hyperspectral image classification method based on deep learning for small-sized samples, a classification model based on multi-scale residual network is proposed. By adding a branch structure into the residual module, the model constructs extraction modules based on spectral features and spatial features, respectively, realizes the multi-scale extraction and fusion of spatial and spectral features, and makes full use of the rich spatial and spectral information in hyperspectral images. In addition, dynamic learning rate, batch normalization, and Dropout are used in the proposed model to improve computation efficiency and prevent overfitting. Experimental results show that the proposed method achieves 99.07% and 99.96% of the overall classification accuracy on the datasets of Indian Pines and Pavia University, respectively. Compared with support vector machines and existing deep learning methods, the proposed model effectively improves the classification performance of small-sized sample hyperspectral images.

Land use change detection is crucial in land resource management and monitoring. Herein, GF-1 multi-spectral remote sensing images, which were captured in Nanchang city of Jiangxi province in 2013—2017, are classified using the image classification method based on support vector machine. Moreover, land use change maps of the study area in those five years are generated, and land use change characteristics are analyzed. Results show that the types of land in the study area are mainly forest land, grassland, water, and building land. In the five years, the vegetation changed the most with a decrease of 54.74 km 2, followed by the water area with an increase of 22.12 km 2, then land building area with an increase of 19.45 km 2, and bare land area with an increase of 5.17 km 2.

Ptychographic iterative and Fourier ptychographic imaging techniques can enhance field of view (FOV) and resolution. The parameter-changed computational imaging technique based on multi-disntance/ multi-height axial scanning and thin cylinder rotation can be used with phase retrieval algorithms to reconstruct the complex-valued fields of objects with high resolution. These imaging techniques possess prominent advantages for enlarging FOV and resolution. We review the recent research progress of the parameter-changed optical coherent diffraction imaging systems in the field of computational imaging. As a kind of indirect imaging tools, the multi-parameter imaging technique combines diffraction imaging with algorithms, which can be used to realize the accurate multi-dimensional characterization of a complex-valued light field.

The sea-sky line is an important feature of sea and air background images, and its detection plays an important role in the division of sky and sea areas and target detection. However, sea clutter, clouds, strong reflection, and dynamic weather conditions in the complex sea and air background environment make it difficult to detect the sea-sky line. This paper aims to address the environmental adaptability problem of sea-sky line detection in a complex background. Three aspects were analyzed in detail: the characteristics of the original image obtained by the visual sensor and those of the sea-sky line in the image; interference factors and their suppression methods; existing methods of generating the sea-sky line. Furthermore, their advantages, disadvantages, and applicable scenarios were elaborated through comparative experiments, and the proposed morphological methods were experimentally compared. Finally, the future challenges and research directions of sea-sky line detection were discussed.

The main task of person re-identification is to use computer vision to match and retrieve specific person across view fields. Compared with the traditional algorithm, deep learning is a more appropriate representative method for the discrimination between persons using data-driven extraction features. This study summarized the background and research history, main challenges, main methods, datasets, and evaluation index of person re-identification. The algorithms of person re-identification were mainly analyzed based on three aspects: feature expression, local features, and generative adversarial networks. The accuracy of 9 common datasets, 3 evaluation criteria, and 14 typical methods of person re-identification on the Market1501 dataset was listed. Finally, the prospects for the future research direction of person re-identification were established.