Aiming at the difficulty of extracting the key contour features of printed circuit boards, an algorithm for transforming the folded edge into the boundary and extracting the key contour feature points is proposed. First, the algorithm establishes a topological structure of the original point cloud data of the printed circuit board by using k dimensional-tree, and realize fast search of the closest k neighborhood points. Pass-through filtering algorithm is used to complete the pre-processing of the printed circuit board point cloud. Second, Random Sample Consensus algorithm is used to extract the plane features with the largest area in the printed circuit board separately, so that the key contour features are spatially separated. The point clustering of the fold edge feature is completed by Euclidean clustering based on normal angle and realize the idea of transforming the folded edge into the boundary. Finally, according to the relationship between set threshold and vector angle between k neighborhood points, one can determine whether the query point belongs to the boundary contour feature point. Experimental results show that the proposed algorithm can extract the key contour feature line of printed circuit board point cloud more completely.

This study proposes a conditional generative adversarial network that improves the performance of underwater image enhancement of different colors. The network adds residual module in residual dense blocks into the generative model, and its dense cascade and residual connections extract image features and ease the gradient disappearance problem. By adding two new loss functions to the objective function, a new network model is established which can make the content and structure of the enhanced images be consistent with that of the input images. The experimental results show that the proposed method has better enhancement performance and visual effect than existing algorithms.

As an open source object detection network, YOLOv3 has obvious advantages in speed and accuracy compared with the object detection network of the same period. Because YOLOv3 adopts a new type of full convolutional network (FCN), feature pyramid network (FPN), and residual network (ResNet), it requires high hardware configuration, leading to high development cost, which is not conducive to the popularization of industrial applications. Therefore, YOLOv3tiny is generally used for detection on embedded platforms. Although the calculation amount is small, the detection effect is far less than YOLOv3. To solve the problem of low detection speed of YOLOv3 on embedded platforms, a simplified version of the network based on YOLOv3 is proposed. Unlike YOLOv3, FCN, FPN, and ResNet, which are helpful for feature extraction, are retained as much as possible. the number of parameters and residual years of each layer is recued, and attempts are made to join densely connected networks and spatial pyramid pooling. Experimental results show that the number of parameters and detection speed of this network is much better than YOLOv3, and the mean average precision is a significant improvement compared to YOLOv3tiny in terms of in the PASCAL VOC2007 and 2012 datasets.

In order to improve the accuracy of multi-spectral image semantic segmentation, a neural network model based on local binary pattern (LBP) feature enhancement is proposed. The model obtains two feature maps from a single infrared image by two LBP feature extraction operators with the size of 3×3 and 5×5, respectively. The RGB image, the infrared image, and the LBP feature maps are imported into a neural network model with a 34-layer residual network for semantic segmentation. The experimental results show that the proposed neural network model can achieve an average accuracy of 60.7% and an average intersection over union of 51.9% on the RGB-Thermal dataset. The results are superior to other comparative methods. At the same time, in the visualization results, the results of proposed model are also more clear and accurate.

Gamut mapping is a key technology for color image transmission and reproduction in different devices, and it is also the core part of modern color management system. However, there are few studies on the quality evaluation of gamut mapping images, therefore, in this paper, a no-reference quality evaluation algorithm based on natural scene statistics for gamut mapping images is proposed. First, the gamut mapping images are converted to the Spatial-CIELAB color space and the three attributes (e.g., luminance, chroma and hue) are extracted. Next, luminance components are decomposed by using Log-Gabor filter, and statistical features are extracted in the frequency domain to characterize image structure distortion and contrast distortion. For the two components of chroma and hue, statistical features are extracted in the spatial domain to characterize color distortion. Then, combined with subjective scores and extracted features, the backward propagation neural network is used to train the image quality prediction model. Finally, this model is employed to assess the image quality. The experimental results prove that the proposed method is superior to the existing no-reference quality evaluation algorithms.

In order to solve the problem that the non-local mean (NLM) algorithm is not accurate enough in measuring the similarity of neighborhood blocks, an improved NLM algorithm based on difference hash algorithm and Hamming distance is proposed. Traditional algorithm measures the similarity between neighborhood blocks by Euclidean distance, and the ability to maintain edges and details is weak, resulting in blurred and distorted images after filtering. Therefore, the difference hash algorithm containing the gradient information is introduced to improve the Euclidean distance, and the Hamming distance of the difference hash value is calculated to measure the similarity of the neighborhood block. Experimental results show that this method can better maintain the edges of details well while denoising, and compared with other algorithms, the running speed of the algorithm is also greatly improved, which has certain application value.

Aiming at the shortcomings of the existing dehazing algorithms, such as transmittance over-estimation, sky color distortion, and poor real-time, a fast and efficient real-time video dehazing algorithm based on pyramid model is proposed. First, pyramid down-sampling is used to obtain the reduced image. Pseudo dehazing image and dark channel confidence are introduced as correction factors to obtain the coarse transmission of the reduced image. Second, the reduced image is restored to original size and refined by a joint bilateral filtering. Finally, atmospheric scattering model and the inter-frame video dehazing theory are combined to restore the degraded video. Experiment results show that, this method can completely dehaze on a variety of scenes. Compared with other algorithms, the improvement of peak signal-to-noise ratio and average structure similarity of this algorithm are 20.153% and 14.056%, respectively. The proposed method is fast, stable, and suitable for real-time video dehazing.

Aiming at the problem of large computation and complicated process of segmentation for whole tumor lesion in segmented magnetic resonance imaging (MRI) three-dimensional images, a fully automatic segmentation algorithm based on deep learning is proposed. A dual pathway three-dimensional convolutional neural network model is constructed on the dilated convolution path filled with jagged holes to extract multi-scale image blocks for training and capture large-scale spatial information. The shallow features are superimposed to the end of the network by using the identity mapping feature of dense connection. The swollen area, enhanced area, core area, and cystic area are segmented in the multi-modal MRI image. The model is segmented and tested in the BraTS 2018 dataset. The results show that the average Dice coefficients of the whole tumor area, core area and enhanced tumor area segmented by the model are about 0.90, 0.73 and 0.71, respectively, which is equal to the performance of the current algorithms and has a high degree of automation integration.

This study proposes an underwater image restoration algorithm based on background light estimation and transmittance optimization to address the problem of color distortion and blur of underwater image. The proposed algorithm defines the transmittance as direct and backscatter component transmittances, which effectively improve the integrity of the imaging model. First, the red attenuation component is compensated, and the histogram distribution range is reset to achieve color balance. Second, the best background light point is selected based on the brightness, gradient discrimination, and hue judgment. The transmittance of the backscatter component is then obtained through a red dark channel prior, and the direct component transmittance is obtained through a nondegenerate pixel point. Finally, the two transmittances and the background light are substituted into the imaging model to obtain a restored image. Experimental results show that the proposed algorithm can effectively balance the chroma, saturation, and sharpness of the image, and the visual effect is close to the image in the natural scene.

To address the limitations of character image generation models, such as ambiguity and lack of texture, this study proposes a pose-guided character image generation model incorporating a fusion feature feedback mechanism. Generative adversarial neural networks are used for training the proposed model. Further, the proposed model is generated during the postural integration and image refinement stages. A fusion feature information feedback mechanism is proposed based on the model to ensure that each stage of the generated model will be subjected to feature comparison adjustment. Inspired by transfer learning, the pre-trained weights of the ImageNet dataset are used as the initial weights of the model feature layer. Moreover, to enhance the robustness of the image generation model and improve the quality of the generated images, corresponding fine-tuning is performed during the training process. Experimental results reveal that the proposed model can obtain more realistic and delicate images of humans that are consistent with human visual perception.

The existing handwritten numerical string recognition algorithm based on over-segmentation is highly complex, and the existing unsegmented recognition algorithm cannot recognize character strings of four digits or more and has a low accuracy rate and the low accuracy. To address these issues, an unsegmented recognition algorithm for handwritten numerical strings based on mask region convolution neural network (Mask-RCNN) is proposed. Because of Mask-RCNN adds parallel full-convolution split-segmentation subnets, it can simultaneously achieve mask segmentation of single digit in the sticky handwritten numerical string and classify digit categories. Results of the test set indicate that after the training of 1-6 numerical strings of images in NIST SD19 dataset and self-built mask-training dataset, the recognition accuracy of the network for character strings of 3 digits, 4 digits and 5 digits is 1.2 percentage, 0.6 percentage and 0.4 percentage higher, respectively, compared with the latest algorithms. The proposed algorithm exhibits significant advantages in recognizing handwritten numerical strings with unrestricted digits and has broad application prospects.

An improved bilinear convolutional neural network (B-CNN) model is proposed to solve the problem of fine-grained classification of crack images of sleeper block shoulder. Using this model, the global information in the image features of the global average pooling link is first used to capture the width information of the fine crack. Then, the fusion of different levels is performed to enhance the ability of feature expression to obtain effective width features and fine-grained classification. Experimental results show that compared with the B-CNN model, the classification accuracy of this model improves by 2 percentage. In terms of the false negative rate, the normal category reduces by 2.3 percentage, and the obvious crack category reduces by 4.55 percentage. Compared with the baseline VGG-D (Visual Geometry Group Network-D) model (6.11 percentage classification accuracy), the normal false negative rate reduces by 7.39 percentage, and obvious crack category reduces by 8.39 percentage. Furthermore, the feature extraction rate for the original is 18.51%, whereas that of our proposed model is 45.31%, which shows that the proposed model can satisfy the need for rapid and accurate imaging of the shoulder for double block-type sleeper crack image classification to meet engineering requirements.

To solve the limitation of the feature fusion target tracking (Staple) algorithm when using the fixed weight fusion method in complex scenes, this study proposes an adaptive-feature fusion and channel weighted anti-occlusion related filtering algorithms with an improved channel confidence. First, to introduce a multi-dimensional feature description, we calculate the channel weights according to the response peak of the filter template on each channel. Then, calculate the reliability of the model based on the response results of the feature model, determine the fusion weight of the model, and complete the feature fusion from the perspective of the response results. Finally, based on the average peak correlation energy of the historical frame and the mean square error of the current and previous frame images, we determine the occlusion of the target and update the model. Comparative experiments with the Staple algorithm are conducted on the OTB-2013 and OTB-100 datasets and the proposed algorithm suggests an improved success rate and accuracy and performs better with respect to many challenging attributes.

In this study, a method is proposed to enhance the low-illumination remote sensing images based on a conditional generation adversarial network, so as to improve their visibility. First, low-illumination images were synthesized as training samples based on the clear images with normal illumination to solve the problem of insufficient sample data. Then, the original low-illumination remote sensing images were converted from the RGB color space to the HSI color space. Subsequently, channel splitting was performed to effectively separate the H, S, and I components, keeping the hue component H unchanged. Further, the conditional generation adversarial network and the improved logarithmic transformation method were used for processing the luminance component I and the saturation component S, respectively. Finally, channel merging was performed to implement the conversion of processed images from HSI color space to the RGB color space. The phenomenon of highly imbalanced sample proportion can be solved by adding focus loss function to the loss function. The experimental results show that the proposed method effectively improves the brightness and contrast of the low-illumination remote sensing images. Furthermore, this study provides novel concepts with respect to the development of low-illumination remote sensing image enhancement methods.

In this paper, a generative adversarial network GI-GAN that combines group interaction information with individual motion information is proposed. First, BiLSTM in the coding layer was used to extract the movement behavior of all pedestrians during the observation period. Second, based on a dual attention module, individual motion information and group interaction information having a high correlation with trajectory generation were calculated. Finally, using the generative adversarial network structure, global joint training was performed and the backpropagation error was obtained. Then, reasonable network parameters for each layer were obtained. Subsequently, the decoder used the acquired context information to generate multiple reasonable prediction trajectories. Experiment results show that compared with the S-GAN model, the average displacement error and absolute displacement error of the GI-GAN model are reduced by 8.8% and 9.2%, respectively, and the predicted trajectories have a higher accuracy and reasonable diversity.

In this study, the hyperspectral imaging technology is employed for accurately and efficiently detecting the surface damage of Korla fragrant pears. Eighty fragrant pears were considered in this study. The hyperspectral images of the intact and damaged samples in the wavelength range of 400-1000 nm were obtained. The hyperspectral image obtained at 863 nm was selected to achieve image mask using the statistical analysis method. The dimension of hyperspectral data was reduced via principle component analysis. Subsequently, the second principle component image exhibiting the most considerable difference between the damaged and background areas was selected to compare with the fourth principle component image via the ratio method of image processing for enhancing the difference between the damaged area and the background area. Finally, the threshold segmentation and morphological operations were used to obtain the damaged areas on the surface of fragrant pears. Results denote that the proposed method can effectively identify the surface damage of fragrant pears. Furthermore, the accuracy, precision, and recall rate of the proposed method are 93.75%, 87.50%, and 100%, respectively.

Aiming at the shortcomings of existing methods for brain tumor image segmentation, this paper proposes a brain tumor image segmentation algorithm based on an improved convolutional neural network. First, DenseNet and U-net network structures are combined to improve the extraction capability for image features. Second, in order to expand the receptive field of the convolution kernel, the cavity convolution is adopted. Moreover, the segmentation results are further finely segmented and output by a fully connected conditional random field recurrent neural networks, thereby obtaining an accurate brain tumor segmentation region. Experimental results show that compared with traditional deep learning methods, the proposed algorithm has an average Dice up to 91.64%, and has a better improvement in accuracy.

Aiming at the problems of low image contrast, color imbalance, and noise in low-light conditions, a low-light image enhancement model based on multi-branch all convolutional neural network (MBACNN) is proposed. The model is an end-to-end model, including feature extraction module (FEM), enhancement module (EM), fusion module (FM), and noise extraction module (NEM). By training the synthesized low-light and high-definition image sample, the model parameters are continuously adjusted according to the loss value of the verification set to obtain the optimal model, and then the synthetic low-light image and the real low-light image are tested. Experimental results show that compared with traditional image enhancement algorithms, the proposed model can effectively improve image contrast, adjust color imbalance, and remove noise. Both subjective visual and objective image quality evaluation indicators are further improved.

In this paper, a new method based on fusing gradient magnitude (GM) and gradient occurrence (GO) is proposed to construct two new descriptors with high resolution. The suppression normalization descriptor constructs a new descriptor by segmenting the GM and inhibiting the smaller GM. GM-GO fusion descriptor is a new descriptor that fuses GM and GO by suppressing partial GO values (which correspond to a smaller GM) to improve the discriminability of the descriptor. Experimental results show that the two proposed methods have higher matching accuracy under the influences of noise and illumination.

In order to further improve the accuracy of breast tumor segmentation, a breast tumor segmentation model based on simple linear iterative clustering (SLIC) and grandient vector flow (GVF) Snake algorithm was proposed. The model first preprocesses the image to reduce redundant information and improve subsequent segmentation efficiency. Secondly, an adaptive value method is proposed based on the texture features of the image, and the image is roughly segmented by SLIC algorithm to describe the initial contour of the breast mass. Finally, the GVF Snake algorithm is used to increase the capture range of the contour edge information, and the segmentation result is obtained by fine segmentation. Experimental results show that the segmentation model can effectively improve the segmentation efficiency and accuracy, which is better than the traditional segmentation algorithm to some extent, and the ideal segmentation results are obtained.

Aiming at the problem of poor visual effect and low image quality of existing low-light images, a low-light image enhancement algorithm based on cascaded residual generative adversarial network is proposed. The algorithm uses constructed cascaded residual convolutional neural network as generator network and improved PatchGAN as discriminator network. First, training samples are synthesized through normal-light image on the basis of Retinex theory. Then, low-light images are converted from red-green-blue space to hue-saturation-value color space. Meanwhile, keeping hue and saturation unchanged, the value component is enhanced through the cascaded residual generator network. Besides, low-light image is enhanced through the way of discriminator network supervising generator network. They struggle against each other to promote the capability of generator network to enhance the low-light image. Experimental results show that the proposed enhancement algorithm obtains better visual effects and contrast in terms of synthetic low-light images and natural low-light images. Especially, for the synthetic low-light images, the proposed algorithm is obviously superior to other comparison algorithms in terms of peak signal-to-noise ratio and structural similarity.

Taking point cloud data as research object, this paper proposes a hybrid octree mixing point cloud index structure which combines a K-dimensional tree (KD-tree) spatial segmentation idea, and realizes efficient management of mass point cloud. In this paper, the space of the point cloud is first divided by the KD-tree idea. On this basis, octree is used for further segmentation to establish an octree-like index structure. Then, in order to achieve better spatial management and neighborhood search, the traditional linear octree coding is improved and optimized. Finally, using five groups of incremented point cloud set as test data, experimental results and comparison analysis show that the octree can make the overall structure of the data organization more reasonable, effectively improve access efficiency, and reduce the memory space. The index structure not only improves the speed of the traditional KD tree construction index but also improves the problem that the traditional octree takes too much space and the neighborhood search takes too long. It achieves reasonable management of massive point cloud space.

Face expressions are affected by factors such as poses, object occlusion, lighting changes, race, gender, and age. Convolutional neural networks are required to learn features more effectively and accurately. AlexNet has low accuracy in expression recognition and strong input image size limitation. In response to these problems, this paper proposes an improved facial expression recognition algorithm for improved AlexNet networks. Introducing multi-scale convolution to the AlexNet network is more suitable for small-scale expression images, extracting feature information of different scales, and cross-connecting feature fusion with higher-level feature information can be realized while the multiple lower-level feature information is transfered downwards, which can reflect the image information more completely and accurately, and construct a more accurate classifier. Because cross-connections will generate parameter expansion, making network training difficult and affecting recognition results. Therefore, we use global average pooling to reduce the dimensionality of low-level feature information, reduce parameters generated by cross-connections, and reduce overfitting. The accuracy of our algorithm on CK+ and JAFFE databases is 94.25% and 93.02%, respectively.

Rain and fog weather seriously affects the quality of outdoor images. In this paper, a new method of defogging based on manifold particle filtering is proposed to solve the problem of edge artifacts. By optimizing the atmospheric transmissivity, the accurate transmissivity is obtained, and the problem of edge artifacts in the depth of field is solved. Aiming at rain marks and unclear problems in removing rain and fog, this paper proposes a method that optimizes the attentive generative adversarial network. By combining the Gaussian model with the generative adversarial network, the background interference is removed, and the accuracy of separation of the background layer from the rain line is improved. At the same time, the manifold particle filter fog removal module is added to the generative adversarial network to recover the clear image without rain and fog. The rain and fog images in the natural scene are used for testing, and qualitative and quantitative analyses are conducted. Experimental results show that compared with the existing rain-removal algorithm, the proposed algorithm can remove the rain line in image effectively, and the details are more abundant. At the same time, the addition of the fog removal module significantly improves the image clarity and the objective index.

Bottled mineral water must be tested to determine the presence of suspended particles prior to being dispatched from the factory. At present, manual methods are used toward this end. These methods are time-consuming and laborious, relying on artificial subjective feelings, and their detection results are not satisfactory. Aiming at this problem, an automatic detection method for suspended particles in bottled mineral water based on computer-vision technology is presented in this paper, which includes image acquisition, recognition of suspended particles, quantity statistics, size parameter detection, and other image analysis processing. On this basis, we develop an automatic detection device for suspended matter in bottled mineral water, describing the structure and working principle of the device, and accomplish the inspection test of suspended matter in bottled mineral water. Results illustrate that the proposed method can both qualitatively and quantitatively detect the number and size of suspended particles in bottled mineral water. The quantitative statistics of suspended particles is accurate. The maximum testing error for the size of suspended particles is 0.28 mm, and the relative errors are less than 6.8%. The proposed device and method could be applied to the detection of bottled mineral water prior to its dispatch from the factory with the characteristic features of accurate detection, saving labor, improving work efficiency, and easy operation.

This paper proposes a feature selection method for the synthetic aperture radar (SAR) target recognition problem based on multi-feature decision fusion that leverages the correlation between sparse coefficient vectors. In the proposed method, sparse representation-based classification (SRC) was applied to solve the coefficient vectors of the individual features, and their correlation was defined. Accordingly, the best combination of features was obtained from the mutual correlation matrix and calculation of the nonlinear correlation information entropy. By investigating the stable intrinsic correlation between the selected features using a joint sparse representation, the target label was determined from the reconstruction errors. Experiments were performed under the standard operating condition, configuration variance, and depression angle variance based on the MSTAR dataset. The average recognition rates of the proposed method for these scenarios reached 99.23%, 96.86%, and 97.46% (30° depression angle) and 74.64% (45° depression angle). A comparison with three existing SAR target recognition methods further validated the effectiveness and robustness of the proposed method.

Herein, U-Net structure was improved to segment magnetic resonance (MR) images of brain tumors to address the loss of information in image segmentation in the full convolutional neural network and low segmentation accuracy caused by fixed weights. Based on the attention module in the U-Net contraction path, the weights were distributed to different size convolutional layers, which is beneficial to information usage for image space and context. Replacing the original convolution layer with the residual compact module can extract more features and promote network convergence. The brain tumor MR image database provided by BraTS (The Brain Tumor Image Segmentation Challenge) is used to validate the proposed new model and evaluate the segmentation effect using the Dice score. The accuracy of 0.9056, 0.7982, and 0.7861 was obtained in the total tumor region, core tumor region, and tumor enhancement, respectively, demonstrating that the proposed U-Net structure can enhance the accuracy and efficiency of MR image segmentation.

Accurately extracting the crack characteristics using the traditional crack detection algorithm is challenging owing to the uneven light intensity, complex background, and significant noise interference of concrete pavement. Herein, to improve crack detection accuracy and reduce computational redundancy, a pavement crack detection algorithm was proposed that used an interlaced low-rank group convolution hybrid deep network combined with low-rank kernel and group convolution. First, crack image datasets were established using the overlapping sliding window clipping method. A robust classifier was generated on the training set to classify crack and no-crack images. Then, the adaptive threshold method was used to obtain a crack binary image with clear edge contours. Moreover, the central axis method was used to achieve the maximum width of the crack. The performance of the model was verified on the testing set. Experimental results show that the detection accuracy is 0.9726, thus showing an improvement over the traditional crack detection algorithm. Compared with the convolutional neural network and its variants, the proposed model involved a significantly reduced set of parameters. Images were processed at 14 frames per second, and good detection results were achieved on three public datasets. For crack widths greater than 2.5 mm, the relative error of the calculation is less than 0.02, which complies with practical engineering requirements.

Utilizing orthogonal matching pursuit algorithm for compressed sensing reconstruction of hyperspectral images is to find the optimal atoms to linearly represent the original signal, so that the residual is continuously reduced to obtain the reconstructed signal. When dealing with the redundant dictionary-based reconstruction, the time consuming mainly exists in its atom matching process and residual updating process, resulting in high computational complexity and difficulty of real-time processing. Aiming at this defect, a Hermitian compressed sensing reconstruction algorithm for hyperspectral images is proposed. The main idea is Hermitian inversion lemma is used to optimize the iterative process of the residual update to improve the execution efficiency of the algorithm. In addition, the artificial fish swarm algorithm is used to find the optimal atoms and accelerate the matching process to further improve the reconstruction efficiency. The experimental results carried out on hyperspectral images show that the computational efficiency of the proposed algorithm can be improved by about 10 times compared with the traditional orthogonal matching pursuit algorithm under the condition of ensuring the reconstruction accuracy.

Understanding the variation of optical properties of polarized light during motion transmission in turbid media is the basis for polarized imaging research on targets in turbid media. In this paper, the degree of polarization of the polarized light used in the experiment was measured first, and the error between the polarization of the polarized light and the ideal polarized light was calculated. Then the particle size and distribution of the fat emulsion particles were measured by the Malvern particle size analyzer. The scattering types of photon and fat emulsion particles were analyzed. Finally, the depolarization of forward scattered light and backscattered light produced by linearly polarized light and circularly polarized light after passing through different concentrations of turbid medium were studied. The results show that the polarization degree of forward scattered light generated in the same concentration of fat emulsion solution is higher than that of backscattered light, and the degree of polarization of both forward scattered light and the backscattered light decreases continuously with the increase of the concentration of fat emulsion solution, and finally stabilizes.

To improve the registration accuracy of a point cloud, the problem of poor robustness of the iterative closest point (ICP) algorithm under the condition of noise interference and data loss caused by a single feature needs to be solved. Accordingly, a point cloud registration method based on weighting neighborhood surface deformation information is proposed. First, to simplify the neighborhood information of points, a neighborhood construction method based on the number of neighboring points as the constraint is proposed, and considering the influence of neighbors on the sampling points, a weighting method is introduced to improve the extraction efficiency of the intrinsic shape signature (ISS) feature point extraction algorithm. Second, the mean value of the normal vector inner product of the neighborhood is calculated to perform the second feature point extraction of the point cloud. Then, the fast point feature histogram (FPFH) is used to describe the feature, and the double constraint condition is used to determine the matching point pair relationship. Finally, in the registration phase, accurate registration is achieved by using the bidirectional k-tree ICP (DTICP) algorithm. Experiment results reveal that the proposed algorithm can effectively register missing point clouds in a noisy environment with better robustness and anti-interference compared with the classical ICP algorithm.

Aiming at the problems that traditional spectral polarization imaging technology requires dynamic modulation, low luminous flux and limited spectral resolution, a new imaging method based on computational spectral imaging technology and pixel-level polarization detection is proposed. The dual-channel format is used for obtaining high-resolution spatial, spectral, and polarization target information through single imaging. Further, a dual-channel experimental device with a coded-aperture spectral polarization imaging channel and a polarization imaging channel is established to obtain spectral data cubes with four polarization states in 25 bands in the range of 450-650 nm, as well as the polarization degree and polarization angle of each band. The spectral resolution of proposed method is better than 10 nm, and the spectral reconstruction accuracy is approximately 86.3%. Furthermore, the spectral reconstruction accuracy is observed to improve by 10.5 percentage points when compared with that of the single-channel imaging method.

In view of the high complexity of artificial feature extraction in traditional machine learning and the low recognition rate caused by inadequate feature extraction in single convolutional network, a new facial expression recognition method based on ensemble convolutional neural network is proposed. The method is to construct an ensemble network (EnsembleNet) model based on integrating an improved VGGNet-19GP model after VGGNet-19 with a ResNet-18 model. The model first trains a single model on the training set to make the single model reach the optimal experiment. Then the ensemble test is performed on the testing set. The average accuracy of 73.854% and 97.611% are obtained on FER2013 and CK+ datasets, respectively. By comparison with the VGGNet-19GP and ResNet-18 models and other existing methods, it is shown that the ensemble-based facial expression classification method has the advantages of more accurate classification and stronger generalization ability.

In this paper, a sub pixel center extraction method is proposed based on dual frequency curve fitting. First, twice threshold method is used to automatically segment light stripe region, and gray-gravity method is used to obtain the initial position of the light stripe region. Then, a normal moving linear fitting is designed to obtain the local normal of light strip curves. Finally, the sub-pixel center coordinate of the light strip is extracted precisely through dual frequency curve fitting of pixel data in the normal direction. Experimental results show that this method can partially eliminate the influence of high-frequency noise. Compared with Steger method, the root mean square error of the strip center extracted by this method is less than 0.1 pixel, and the speed is about 26 times of it. This method can be applied to the extraction of light stripe center of various materials, which provides a reference for the extraction of laser stripe center in industrial applications.

On the issue about optimizing the indices of object detection for YOLOv3 model in traffic scenes, we make the model indices further optimized by combining various strategies and tricks in the process of training, meanwhile, and we propose an improved anti-occlusion strategy based on Cutout. The optimization does not involve changes of the original YOLOv3 network structure, and there is no impact on FPS after optimizing. Comparison experiments are conducted on both PASCAL VOC and KITTI 2D, the obtained results show that these strategies and tricks can significantly improve the performance of YOLOv3 model. Full code has been released, click to view or download at: https://github.com/LiweiDai/YOLOv3-training-optimization-with-applying-ACDC.

In micro-expression recognition, directly using the original micro-expression sequence achieves sub-satisfactory results, and the existing algorithms often employ a single feature map rather than fusing multiple feature maps. To address these problems, this paper proposes a new micro-expression recognition algorithm that fuses motion feature maps after extracting the features to obtain more accurate recognition results. The proposed algorithm uses the fused deep learning framework between convolutional neural network (CNN) and long-and-short memory (LSTM) network. Different algorithms are evaluated on the CASMEII micro-expression database. Experimental results show that the proposed method performs better compared with other algorithms.

Quickly and efficiently distinguishing and eliminating weeds is one of the keys to improving the extraction accuracy of pitaya plants. In this study, a high-resolution aerial image is acquired using a four-rotor unmanned aerial vehicle (UAV) platform with a visible light lens. The spectral characteristics of pitaya plants and weeds in R, G and B channels are then analyzed, and the close color difference vegetation index (CCVI) is constructed based on the pixel digital number (DN) values. Through OTSU threshold segmentation, majority/minority analysis and cluster hole filling, the mainstream indices including VDVI, EXG, and NGRDI are compared with CCVI. Results show the following: 1) for the pitaya plant plot with a high or full coverage rate for weed, the CCVI extraction effect is better, whereas the other three indices have poor classification effect in the plot where weeds and plants coexist; 2) for the three research ROI samples, the overall average accuracy and the Kappa coefficient are 94.60%, 0.9417, respectively, and for the test sample, the overall extraction accuracy and the Kappa coefficient are 94.33% and 0.9328, respectively. Thus, it is verified that the extraction accuracy of plants with similar habitats in different regions is fairly similar. Results confirm that the CCVI can be used to identify and extract the individual pitaya plants from the weeds with the UAV remote sensing scheme, and its extraction effect is good. The proposed method can be applied in conjunction with VDVI, EXG, and NGRDI.

Nano-computed tomography (nano-CT) is of great significance for the development of science and technology and the progress of production technology, especially in frontier researches of biomedicine and new material properties. The principle and classification of nano-CT imaging, as well as the implementation, advantages, and disadvantages of the nano-CT system, are reviewed. As for the specified technology detail, the principles, properties, manufacturing bottleneck, and application status of optical components in nano-CT are described. On this basis, the existing problems in nano-CT imaging system and its applications are proposed, and the future development of nano-CT imaging is summarized.