Search

NeRF is a deep learning model aimed at modeling three-dimensional implicit spaces, and it holds significant value in the representation and rendering of 3D scenes. However, due to the complex training process, substantial computational resources, and time requirements, the usability and practicality of the NeRF algorithm are somewhat limited. Addressing the pain points of NeRF optimization has become a hot topic in the field of computer vision. This paper aimed to provide a comprehensive review of the optimization and application of NeRF. Firstly, it delved into the basic principles of NeRF and outlined the current optimization status from the perspectives of rendering quality, computational complexity, and pose. Secondly, it enumerated the application scenarios of NeRF to provide references for future, more efficient and practical algorithmic optimizations. Finally, it summarized the strengths and limitations of NeRF and proposed potential future directions tailored to harness the tremendous potential of NeRF in 3D rendering, scene synthesis, and beyond. [ABSTRACT FROM AUTHOR]

Aiming at the problem that feature extraction is inaccurate and the reconstruction image is not natural enough in blind super-resolution reconstruction, this paper proposed an image blind super-resolution reconstruction based on degradation aware and sequence residuals. This paper proposed a mini-residual group combined degeneration aware and sequence residuals as the backbone network. Then the method constructed a symmetrical enhanced multi-scale residual block. In the image reconstruction part, this paper used the bottleneck attention module and the sub-pixel convolutional module to emphasize the multi-dimensional elements of the image. Finally, the method made a global residual connection. Compared with the current representative algorithm DASR, experiments show that the PSNR and SSIM of the proposed algorithm are improved 0.145dB and 0.0014 on Set14×2, and the PSNR of the proposed algorithm is improved 1.898dB and 0.252dB on Set14×3/4, respectively. The proposed algorithm achieves better performance than several current image super-resolution algorithms on five standard test sets. [ABSTRACT FROM AUTHOR]

Aiming at the problem of insufficient feature extraction ability of deep clustering algorithm on multivariate time series data( MTS), This paper proposes a new deep clustering structure model (MDTC. In order to extract the key features of MTS and realize dimensionality reduction, one-dimensional convolution learns the attribute and temporal dimension feature representation of MTS and the autoencoder structure composed of network layers such as recurrent neural network; To improve the model’s ability to represent temporal features, this paper proposes MCBAM temporal attention module, which is used to enhance the representation features of different time periods in the MTS sequence. This paper conducts experiments on nine publicly available UEA multivariate time series datasets, compared with other autoencoders, the autoencoder structure of the model improves by 2%-9% on 7 data sets; Compared with other attention modules, the MCBAM module of the model improves by 0.3%-2% on the six datasets. Experiments show the effectiveness of the MDTC model structure and MCBAM module, and the model has excellent performance compared with other clustering algorithms. [ABSTRACT FROM AUTHOR]

In order to address the shortcomings of the facial landmarks detection models, which cannot model the relations between long-distance landmarks, this paper proposed a parallel multi-branch architecture combining with convolution and Transformer for facial landmarks tasks, called MCTN, it utilized the dynamic attention mechanism to model the long-distance relations between facial landmarks. The multi-branch parallel structure designing allowed MCTN to include shared weights, global information fusion and other merits. What’s more, this paper proposed the novel Transformer structure, Deformer, which could make the MCTN focused attention weights faster on sparse and meaningful locations and solved the problem of slow convergence of Transformer. MCTN reached 4.33%,3.12% and 3.15% normalized average error respectively on the WFLW,300W and COFW datasets, the results show that MCTN utilizes Transformer with CNN multi-branch parallel structure and Deformer structure to dramatically outperform other facial landmarks localization algorithms based on convolution network. [ABSTRACT FROM AUTHOR]

In recent years, various human action recognition algorithms have achieved excellent recognition performance based on a large number of labeled samples. However, in practical applications, acquiring training samples and their corresponding labels is time-consuming and laborious, which limits the actual implementation of the algorithm. This paper summarized the deep learning algorithms for action recognition under weak supervision and few-shot learning. Firstly, in the case of weak supervision, it classified and summarized the semi-supervised action recognition methods and unsupervised domain adaptation video action recognition methods separately. Then, it reviewed video action recognition algorithms based on few-shot learning in detail. Further, it summarized the relevant human behavior datasets and analyzed and compared the performance of various relevant video action recognition algorithms on these datasets. Finally, the paper summarized the full text and discusses the future development trend of human action recognition. [ABSTRACT FROM AUTHOR]

Aiming at the problem that the traditional deep learning fish recognition method has a low accuracy rate and the parameters cannot be determined adaptively during the model training process. This paper proposes an improved ShuffleNet fish identification method based on Particle Swarm Optimization( PSO) and Transfer Learning(TL). The research takes 20 species of fish as the object, uses the particle swarm algorithm to take the model's loss function as the fitness function, optimizes the two hyperparameters of batch size and learning rate, and uses the transfer learning method for training to construct TL-PSO-ShuffleNet model. The research shows that compared with the models of AlexNet, MobileNet, and ShuffleNet, the recognition accuracy rate is increased by 57. 89%, 30. 43% and 23. 28%, respectively. The fish identification method proposed in this paper has the characteristics of high accuracy and self-adaptive parameter setting, which provides a reference for the research on automatic fish identification. [ABSTRACT FROM AUTHOR]

The purpose of image stitching is to obtain a high-definition, seamless panoramic image. Existing methods rely on the accuracy of feature matching, which will misalign images and produce errors such as artifacts and distortions.This paper proposed a new unsupervised real scene image stitching network that can adapt to real scene stitching in the presence of moving targets, ensuring no loss of accuracy in the panorama. It included two networks of alignment and reconstruction, excluding the negative influence of moving targets and misleading regions on the transformation matrix through content branching, and optimizing image details by constraining the reconstruction process through edge branching to achieve high-definition and artifact-free stitching effects. The experimental results show that the method’s RMSE, PSNR, and SSIM reached 1.81, 26.56, and 0.85, respectively. The objective evaluation indexes are better than other classical algorithms overall, and the user research results also indicate that the method obtains higher definition of panoramic images. The method effectively accomplishes unsupervised image stitching in real scenes and can be generalized to stitching tasks in other scenes. [ABSTRACT FROM AUTHOR]

Real-time target detection in traffic scenarios is the prerequisite of electronic monitoring, automatic driving, and other functions. In view of the low detection efficiency of existing target detection algorithms and the low accuracy of most light target detection algorithms, which are easy to obtain wrong or insufficient target detection, this paper adopts the improved YOLOv5 target detection algorithm for model training, and the pseudo-label strategy for training process optimization. Then, the KITTI traffic target dataset tags are merged into three categories for model training and testing. Through the experimental comparison, the improved YOLOv5 model in this paper achieves 92.5% mAP in all categories, which is 3% higher than the original YOLOv5 training model. Finally, the three categories of the trained models are deployed on the Jetson Nano embedded platform for inference testing, and TensorRT is used to accelerate inference. The average inference time per frame of image is 77 ms, which meets the goal of real-time detection. [ABSTRACT FROM AUTHOR]

In dynamic gesture recognition methods based on deep learning, aiming at the problems of large network scale, simple dimension of extracted spatiotemporal features, and insufficient extraction ability of effective features, this paper proposed a deep network framework. Firstly, this paper applied a novel spatiotemporal convolution module based on multiscale information fusion in the time domain to improve the 3D residual network structure, greatly reduced the size of the network and obtained rich spatiotemporal receptive field characteristics. Then it introduced a spatiotemporal feature channel attention mechanism with global information synchronization, and used a few parameters to construct the global dependency between feature maps, and the module could obtain the key features of dynamic gestures more efficiently. The experimental results on the self-built gesture dataset SHC and the public gesture dataset SKIG show that the proposed gesture recognition method has fewer parameters and more powerful multiscale spatiotemporal feature extraction ability, and gains a higher recognition rate than the current mainstream algorithms. [ABSTRACT FROM AUTHOR]

High-quality learning low-dimensional representation of nodes in the graph is a current research hotspot. The existing shallow model methods cannot capture the nonlinear relationship of the graph structure, and the graph convolution model in the graph neural network technology will cause an over-smoothing problem. At the same time, how to determine the role of different hop number relationships in graph representation learning is also a problem that needs to be solved in the research. To solve the above problems, this paper proposes a deep learning model based on T（T＞1） feedforward neural networks. The framework uses deep learning models to extract the nonlinear relationship of the graph structure, and T sub-models effectively capture the local and global（higher-order） relationship information of the graph, and they give different roles in the final vector representation to take advantage of different hop relations. Experimental results on vertex classification and link prediction tasks show that the framework is competitive with existing methods, the benchmark algorithm can be improved by about 20%. [ABSTRACT FROM AUTHOR]

Traditional manually designed feature descriptors are prone to cause a large number of mismatches in the face of multi temporal remote sensing images with large geometric and radiation differences and obvious changes in ground features. This paper developed a convolutional neural network matching algorithm with double screening and proposed double constraints to extract the deep semantic feature map of the image through CNN, it selected the key points that satisfied both the priority maximum principle and the precise extreme value principle from the feature map, and extracted 512-dimension descriptors from the corresponding positions. In the feature matching stage, the strategy of combining the rough to fine reverse matching constraints and RANSAC constraints ensured that the false matching points were eliminated effectively, and retained a certain number of correct matching points, which improved the matching accuracy. The experimental results show that the number of matches is moderate, the correct matching rate is improved compared with other algorithms selected in the experiment, and the image registration error is reduced by about 2 pixels. The running time is obviously improved, matching effect is good, with certain adaptability and robustness. [ABSTRACT FROM AUTHOR]

Single-view 3D reconstruction based on deep learning is a research hot spot at present. In order to discover more high-frequency details, SDF-SRN algorithm introduces positional encoding, but neural network is easy to overfit without accurate supervision, and reconstructs uneven surface. To solve the problem, this paper proposed the network model based on sparse feature. The model enabled the network that preferred to overfitting to predict high-frequency residual by residual learning. The feature extraction network extracted sparse features and the global features. Then one hypernetwork took the sparse features as input and generated prediction shallow head. This shallow head predicted low-frequency part of signed distance function. Another hypernetwork took global features as input and generated another shallow head. This shallow head predicted high-frequency residual. It fused two predictions of shallow heads into final signed distance function. Spectrum analysis shows that the design purpose of network is achieved. Compared with other smooth surface reconstruction schemes, the network can achieve smoother surface reconstruction with enough details. It overcomes the overfitting of SDF-SRN. The qualitative and quantitative comparison with other advanced single-view reconstruction approaches show the superiority of the proposed approach. [ABSTRACT FROM AUTHOR]

The image rain removal algorithm analyzes and processes rain images to remove rain streaks and restore clean background scenes, which helps to improve the recognition accuracy of computer vision task, so that it becomes a hot spot in current research. In order to systematically understand the current research status and development trends in this field. Firstly, this paper introduced some typical synthetic rain models. Secondly, it analyzed typical image rain removal algorithm models and methods from two aspects, such as model-driven and data-driven. Then it compared the algorithm evaluation metrics and public rain datasets respectively. Finally, this paper prospected the future development trend of single image rain removal algorithms. [ABSTRACT FROM AUTHOR]

The recommendation algorithm based on graph neural network can extract the association relationship between users and goods. Traditional methods can’t extract this relationship. At present, most of these algorithms ignore the general prefe-rences in the review data of users and products. In order to solve this problem, this paper proposed a new method. This me-thod used the graph neural network to extract association relations, and the advantage of deep learning to extract the general preferences, and carried out feature fusion to improve the recommendation effect. This paper conducted comparative experiments and ablation experiments on four sets of public data sets to verify the effectiveness of the proposed method. The evaluation indexes include the recall rate and normalized discounted cumulative gain. Experiments show that this method is more effective than the existing algorithms. The feature fusion of the two networks can improve the recommendation effect. [ABSTRACT FROM AUTHOR]

Aiming at the problem of fine-grained visual categorization, this paper proposed an effective algorithm to achieve end-to-end fine-grained visual categorization. The ECA module in ECA-Net was a channel attention mechanism with significant performance advantages. It model-fused with the classic network ResNet-50 to form the ResEca. Then, it used the object-level image positioning module and the part-level image generation module to generate object-level and part-level images. Those images combined with original images could be as the input of the new constructed network Tb-ResEca-Net. This paper trained the 95. 1% and 95. 3% respectively on the test set of the corresponding dataset. The experimental results show that this method has higher classification accuracy and stronger robustness compares with other traditional fine-grained classification methods, which is an effective fine-grained image classification method. [ABSTRACT FROM AUTHOR]

Single object tracking( SOT) is a technique to analyze the motion status and provide localization of the single target by using target appearance and context information in video. It has promising applications in intelligent surveillance, intelligent interaction, navigation and guidance, etc. However, problems such as occlusion, background interference and appearance variation led to slow progress in practical applications. With the rapid development of deep learning in recent years, the study of using deep learning techniques to optimize SOT algorithm has become one of the hot spots in computer vision. Around the SOT algorithm based on deep learning, this paper respectively provided outline and analysis of each of the six aspects of correlation filter, Siamese networks, meta-learning, attention, recurrent neural networks and generative adversarial networks according to the core algorithms, after analyzing the basic principles of SOT. In addition, this paper summarized the current state of research and proposed the development trend and optimization ideas of the algorithms. [ABSTRACT FROM AUTHOR]

[Objectives] Weaned piglets are the focus of attention in pig breeding industry. Accurate tracking of piglets in group breeding environment is the basis of analyzing individual behavior and monitoring individual health of piglets. This paper proposed a target tracking method for weaned piglets based on deep learning. [Methods] The detection model of weaned piglets was designed by using centernet detection algorithm based on center point to optimize the detection part of DeepSORT algorithm. The model was trained and optimized for extracting piglet target appearance characteristics during tracking. Combined with Kalman filter and Hungarian matching algorithm, the re-identification was improved to realize the real-time detection and tracking of weaned piglets with deep correlation measurement. [Results] The algorithm test results showed that the mean average precision and recall of weaned piglets were 99.0% and 78.6% respectively. The multi-target tracking accuracy indexes of MOTA (multiple object tracking accuracy) and MOTP (multiple object tracking precision) were 96.8% and 81.8%, respectively. [Conclusions] The method proposed in this paper can improve the tracking difficulty caused by the high similarity of weaned piglets' appearance and adhesion occlusion to some extent. It can accurately track the weaned piglets in the herd environment, and provide technical support for the subsequent individual behavior analysis of piglets. [ABSTRACT FROM AUTHOR]

Abstract: Accurate identification of cotton top bud is important for cotton topping operation to detect cotton terminal bud accurately in field, a recognition method using Faster R-CNN (Faster Region Convolutional Neural Networks, Faster R-CNN) by integrating dynamic mechanism was proposed to solve the recognition difficulties caused by the small size of cotton terminal bud when it is topped in densely planted fields. The RegNetX-6.4GF model was used as the backbone network to improve its image feature extractor capabilities. Due to number of proposals under a higher IOU(Intersection Over Union, IOU) as well as the matching degree between anchor and the target shape affect the performance of the detector, the method proposed in this paper changed the original anchor generation mechanism by combining FPN (Feature Pyramid Network, FPN) and Guided Anchoring in RPN (Region Proposal Network, RPN), which will cause the distribution of the proposals generated by the RPN of the algorithm under different IOUs dynamically change during the training process. To adapt the dynamic change of proposals distribution, we integrated Dynamic Region Convolutional Neural Networks ( Dynamic R-CNN) in Faster R-CNN, which can dynamically adjust the value of IOU to obtain high quality proposals. And the GROIE (Generic ROI Extractor, GROIE) mechanism was inducted to extract ROI (Region of Interest, ROI) to improve the feature fusion capability. In this paper, 4 819 images of gossypium hirsutum population which contain seven leaf types were taken from the top of cotton plant at distance of 30-50 cm (medium distance) and 50-100 cm (long distance) under uniform light, oblique strong light, direct strong light, and shadows. Those images were processed as MS COCO 2017 format dataset and assigned to the training set, validation set, and test set, which contained 2 815, 704, and 1 300 pictures respectively. The experimental results demonstrated that FPS (Frames Per Second, FPS) of proposed model was up to 10.3 frames/s and the Mean Average Precision (MAP) of bud identification reached to 98.1% which was 3.2 percentage points higher than original Faster R-CNN model. The validation set were used to compare performance of mainstream recognition algorithm and proposed method. Results showed that the improved Faster R-CNN's MAP was 7.3% higher than original Faster R-CNN, which was also higher than RetinaNet, Cascade R-CNN (Cascade Region Convolutional Neural Networks, Cascade R-CNN) and RepPoints by 78.9%, 10.1% and 8.3% when IOU was set to 0.5. The improved Faster R-CNN proposed in this paper meets the accuracy and real-time requirements of cotton topping operation. [ABSTRACT FROM AUTHOR]

In order to protect personal information in images, this paper proposed an adversarial image perturbations algorithm to combat deep neural network, which could mine and discover personal image knowledge. It transformed the problem of adversarial example generation into a multi-objective optimization problem with constraints. Considering the classification confidence of the neural network, the location of the perturbed pixels and the chromatic aberration, this paper obtained the adversarial examples iteratively by using the differential evolution algorithm. On MNIST and CIF AR-10 dataset, based on deep neural network LeN et and ResN et, the algorithm generated the experiment of adversarial examples. This paper compared and analyzed the success rate, number of perturbation pixels, optimization effects and spatial characteristics of the adversarial examples. The results show that the proposed algorithm still can effectively combat the deep neural network in the case of few disturbed pixels ( the average number of perturbation pixels is 5) . The algorithm significantly optimizes the location and chromatic aberration of the perturbed pixels, so as to protect personal information without destroying the original image. This study is helpful to balance the relationship between information technology dividend sharing and personal information security, and provides technical support for the research of adversarial examples generation and classification spatial features in deep neural networks. [ABSTRACT FROM AUTHOR]

Deep matrix factorization adopts deep nonlinear mapping to break through the bottleneck of bilinear relationship affecting the performance of the recommendation system in matrix factorization. However, it doesn't take into account users' preference for unrated items, and its recommendation performance isn't advantageous for large-scale data with high sparsity. This paper proposed a recommendation algorithm for fusing matrix completion and deep matrix factorization. First, it filled the unknown elements in the original scoring matrix by the matrix completion model. Then, based on the completed matrix, it constructed the latent vectors of users and items by deep learning model. Finally, this paper tested the proposed model on the MovieLens and SUSHI datasets. Experimental results show that the proposed algorithm significantly improves the performance of the recommendation system compared to deep matrix factorization. [ABSTRACT FROM AUTHOR]

This paper presented an image dehazing method based on deep learning to solve the problem that the remote sensing images had reduced image sharpness due to haze. Firstly, this algorithm distorted the original atmospheric scattering model to get an end-to-end defrosting model. Then, it unified several unknown parameters to one parameter and estimated the unknown parameters by using multiscale convolution neural network. Finally, it brought the parameter estimates into the defrosting model to get a haze-free image. For no reference image dataset, this paper used the existing dataset to preliminary train the network, then added the self-built dataset to secondary train the network. The experimental results show that, compared with the related defrosting algorithms, the proposed algorithm improves the visual effect and objective index to some extent, and effectively improves the clarity of remote sensing images in haze weather conditions. [ABSTRACT FROM AUTHOR]

The community detection method can mine hidden cluster structure in the network, which is of great significance for the analysis of the structure and function of complex network. In recent years, with the explosive growth of network data and the diversity of network evolution, a large number of methods and frameworks for community detection that can handle different scenarios have emerged. In order to deeply understand the research status and development trend, this paper reviewed the community detection methods. This paper firstly classified these algorithms, introduced each kind of algorithm in detail and made analysis and comparison, and then introduced some commonly used evaluation indicators and expounded the application scenarios. Finally, this paper looked forward to the future research directions of this field. [ABSTRACT FROM AUTHOR]

Most of the existing methods only connect multi-modal features, without considering the complex connection between modes. To overcome these limitations, this paper proposes an end-to-end model named factorized bilinear pooling for multi-modal fusion fake news detection. First, the algorithm effectively fuses the text and visual features captured by multi-modal feature extractor utilizing factorized bilinear pooling method, and then cooperates with fake news detector to classify posts as fake or not. The role of event classifier is to remove the event-specific features. Experimental results based on two multimedia datasets, Twitter and Weibo, demonstrate the effectiveness of the proposed algorithm. Experimental results show that the proposed algorithm can effectively fuse multi-modal data, narrow the heterogeneity difference between modes, so as to improve the accuracy of fake news detection. [ABSTRACT FROM AUTHOR]

To solve the problems that the objects with different semantic classes are identified as co-salient objects in current co-saliency detection methods, this paper proposed a CNN and semantic correction-based co-saliency detection method (CSCCD). The proposed method first adopted the guided super pixel filter to process the super pixels obtained by SLIC and the saliency results generated by DSS, which showed clear object boundaries. Then it utilized Mask R-CNN to extract semantic features. It proposed the definitions of image semantic feature and semantic consistency. It also defined the image group semantic correction to solve the problem that detected the objects with different pose belonging to a semantic class as different semantic classes. With the concept, this paper defined image group semantic correlation class, solving semantic correlation problem of multiple images. It generated the final co-saliency detection results by fusing the saliency detection regions with the image group semantic consistent regions. The experimental results on public benchmark datasets show that this algorithm can effectively highlight the whole and outline of the object, and its comprehensive performance in objective quantification is obviously improved. [ABSTRACT FROM AUTHOR]

In order to solve the problem that the existing super-resolution algorithm based on deep learning didn't make full use of the feature information of each level, resulting in low reconstruction accuracy and large parameter quantity, this paper proposed a double dense connection structure named densely channel attention skip connection network. In the inner structure of the network, it improved the original dense cascade block to generate a channel separable dense cascade block. The outer structure adopted a densely residual connection and attention mechanism to fuse the features extracted by the dense block to achieve the goal that less convolution layer and higher precision effect. This paper tested the network models on several benchmark datasets . The results show the proposed model has higher accuracy and fewer parameters than the other models. [ABSTRACT FROM AUTHOR]

In order to understand the development of outlier detection technique detailly, this paper overviewed the outlier detection techniques in recent years. Firstly, it introduced and summarized the definition of outliers, the causes of outlier and classification of the outlier mining algorithms. Next, it overviewed outlier detection methods based on neighbor, based on distributed framework and based on deep learning. Especially, it discussed the most representative methods in outlier detection, and pointed out the merits and drawbacks. In the end, this paper concluded and discussed the future direction of outlier detection techniques. [ABSTRACT FROM AUTHOR]

In view of the shortcomings of traditional salient target detection methods in detecting multiple salient targets at different scales. This paper presents a salient object detection algorithm with deep multiplexing of multi-scale features. The network model consists of vertically stacked bidirectional dense feature aggregation modules and horizontally stacked multiresolution semantic complementary modules. First, the bidirectional dense feature aggregation module extracts semantic features of different resolutions based on the ResNet backbone network, and then performs adaptive fusion on the top-down and bottom-up paths in turn to obtain multi-scale representation features at different levels; The multi-resolution semantic complementation module fuses the multi-scale features of two adjacent levels to eliminate the mutual crosstalk between features at different levels and enhance the consistency of prediction results. The experimental results on 5 benchmark datasets show that the method can achieve the highest Fmax, Sm, MAE of 0.939, 0.921, 0.028, and the detection rate can reach 74.6 fps, which has better detection performance compared with other comparison algorithms. [ABSTRACT FROM AUTHOR]

Salient object detection (SOD ) aims at detecting the significant parts in the images automatically and efficiently, which plays an important role in computer vision tasks such as object detection and recognition, key points location, visual tracking, semantic segmentation and so on. This paper focused on a comprehensive review of salient object detection, which covered the closely related fields, models, datasets, applications and future directions. It introduced the relationship between salient object detection and closely fields at first. Then, it analyzed salient object detection algorithms in terms of flowcharts, ideas, innovations, performance as well as applicability. Thirdly, it introduced the public datasets which improving continuously with the development of the saliency detection models. Finally, it discussed a prospect of future researches and several real application scenes of salient object detection. [ABSTRACT FROM AUTHOR]

Based on video action recognition, this paper firstly reviewed the traditional action recognition algorithms based on RGB, including traditional algorithms and deep learning algorithms. Because of the influence of background lighting, RGB-videobased action recognition is low accuracy, but it has rich color appearance information. Then this paper analyzed and summarized action recognition algorithms based on RGB-D, which was divided into depth sequence, bone and multi-feature fusion. RGB-D video has multiple modes and can provide more information for action recognition, which performs better than RGB video but also brings new challenges. Finally, this paper prospected the common data sets and the possible future development direction. [ABSTRACT FROM AUTHOR]

The existing pig counting system based on image processing are seriously affected by light conditions, and the counting accuracy is poor when pigs are crowded and obscured.To realize the intelligent pig counting, the paper proposed a counting scheme based on the improved instance segmentation algorithm was proposed in this paper. Aiming at the problems of image illumination and target edge blur, Laplace operator was used to preprocess the images. The feature extraction network of MASK R-CNN network was improved by using Resnet-152 as the Mask R-CNN feature extraction network, and the original Feature Pyramid Network(FPN) was followed by a bottom-up enhancement path, which directly fused the low-level edge features with the high-level features to improve the recognition ability of the target edge bluring. The non maximum suppression process and loss function were optimized and improved to improve the segmentation accuracy. The experiments were carried out at three different real pig farms to verify the counting accuracy, respectively in Hebei province, Jilin province and Inner Mongolia. The size of the cage in Hebei pig farm was 5.5 m × 1.8 m, with an average of 12 pigs in a single pen, with a feeding density of 1.21 pigs/m²; the size of the cage in Jilin pig farm was 5.5 m × 3.9 m, with an average of 22 pigs in a single pen, and the average rearing density was 1.03 pigs/m²; the size of the cage in Inner Mongolia pig farm was 11.4 m × 5.28 m, with an average of 80 pigs in a single pen, the average feeding density was 1.32 pigs/m². The RGB camera is was installed on the top of the pen and acquired the image in daytime. 2 400 images were collected in total, and 2000 images were selected after image preprocessing, and 1 250 images of three pig farms were selected as the original data set according to the ratio of 2:2:1. The training set and verification set were enhanced to 1 500 and 150 images and 250 images for the test set, The experimental results showed that 98 images could realize exact counting and 2 images missed 1 pig in the Jilin pig farm, the accuracy of pig counting was 98%. In Hebei pig farm, 99 images could realize the exact counting and the accuracy of pig counting was 99%. For Inner Mongolia pig farms with high feeding density, the accuracy of pig counting was 86%, among the 50 test images, 7 images missed detection, 4 images missed 1 target, 3 images missed 2 targets, The results can provide the application of the artificial intelligent in agriculture field. [ABSTRACT FROM AUTHOR]

Aiming at the knowledge overload in the process of online learning and the problems of data sparsity and cold start in the traditional recommendation algorithm, this paper proposed an improved learning resource recommendation algorithm based on MLP. The algorithm used the advantage of MLP in nonlinear data processing, converted the prediction mode of vector multiplication of learner characteristics and learning resource characteristics into the input mode of MLP, and improved the DN-CBR neural network recommendation model. In order to verify the effectiveness of the model, this paper built a dataset with the online learning platform data of iCourse as the sample. Compared with DN-CBR model, the improved model improves the normalized cumulative loss gain and hit rate by 1. 2 % and 3 % respectively on the dataset, which effectively improves the recommendation performance of the model. [ABSTRACT FROM AUTHOR]

Vehicle detection and identification in urban roads is of great significance for improving traffic safety and developing intelligent transportation. The traditional detection method relies on the features of manual extraction, which has been difficult to apply the complex and variable traffic scenarios, and has the defects of low recognition accuracy and high time complexity. The deep learning model can automatically extract effective features, and the generalization ability is strong, but it is difficult to classify similar vehicles more closely. To this end, this paper proposed a small vehicle target detection algorithm based on residual network. The algorithm changed the connection form of the traditional convolutional neural network to a residual connection mode based on local connection and weight sharing. At the same time, this paper changed the number of network structure control parameters, fused the features of different levels of the picture, applied the pooling layer of the region of interest to normalize the front layer features, and finally obtained the confidence and correction parameters of the target frame through the classification layer and the regression layer. Experiments show that the improved model can enhance the learning ability of the network under the premise of ensuring time efficiency, improve the average accuracy value, and obtain good detection results on the detection of similar small vehicles. [ABSTRACT FROM AUTHOR]

Aiming at the problem of autonomous localization and environmental reconstruction, SLAM based on visual sensor has become a hot research topic at this stage. In order to deeply analyze the current situation of visual SLAM, this paper summarized the related algorithms and results. Firstly, it briefly summarized the concept, characteristics and research significance of visual SLAM. Then, it analyzed the visual odometry algorithm in depth, and described the classical visual odometry algorithms in detail, which included feature-based method, optical flow-based method and direct method, and introduced the landmark achievements of classical visual SLAM algorithm. Secondly, this paper introduced the research progress visual SLAM based on deep learning according to supervised learning and unsupervised learning, and summarized the algorithms. In addition, it analyzed the integration of visual SLAM and inertial navigation. Finally, it proposed the future development trend of visual SLAM. [ABSTRACT FROM AUTHOR]

Aiming at the problems of time-consuming and high error rate in handwriting test scores statistics a handwriting test paper recognition method based on improved convolution neural network algorithm was proposed. In order to simplify the classification of the score» the score column of the handwritten test paper was divided to obtain ten types of numbers from 0 to 9. In order to improve the efficiency of handwritten score recognition, a classification and recognition algorithm of convolution neural network and Bayesian was proposed. The constructed convolution neural network model was used to extract the characteristics of handwritten digits. The PCA algorithm was used to reduce the dimensionality of the features.The hayes classifier was used to distinguish ten kinds of numbers from 0 to 9. the auuracy and efficiency of the algorithm were verified in the MNIST database.The paper score summation model was established and automatic summation is performed after recognition. The experimental results show that for the recognition uf the handwritten scores of 1 188 lest papers in 3 courses♦ the algorithm in this paper has a recognition rate of 98. 23% compared with other algorithms, the average recognition time per test paper is 7. 5 s, and verified its practicality. [ABSTRACT FROM AUTHOR]

When modeling users' preferences, current researches of group recommendation usually ignored the mutual influence between group preference and individual preference and the problem of modeling initialization. To address these issues, this paper proposed a new group recommendation model called ranking based hybrid deep tensor factorization model, namely R-HDTF model. First of all, this paper developed a hybrid deep learning-based initialization method, which utilized deep denoising autoencoder to pretrain the initial values of the parameters for the R-HDTF model. Then, it proposed a pairwise tensor factorization model to capture the correlation among group, individual and item. Finally, it used the Bayesian personalized ranking( BPR) metric to optimize the loss objective function of tensor factorization and learn the parameters of the proposed model. Experimental results on real data sets show that the performance of the proposed algorithm outperforms traditional group recommendation algorithm. [ABSTRACT FROM AUTHOR]

In order to give a comprehensive understanding of SLAM, this paper first gave an overview of the progress of visual SLAM community has made over the last 30 years in this survey. Then it presented the non-linearity of the mathematical model and computational complexity in visual SLAM algorithms. This paper mostly focused on the latest achievements and approached improving the accuracy of pose estimation, building a globally consistency representation of the environment and promoting computation efficiency. It also surveyed the failure modes of current visual SLAM algorithms for large scale, full life cycle implementation and the different ways to address that. Finally, it discussed the potential connections between deep learning architectures and visual SLAM state estimations. [ABSTRACT FROM AUTHOR]

Depth estimation is critical to 3D reconstruction and object location in intelligent agricultural machinery vision system, and a common method in it is stereo matching. Traditional stereo matching method used low-quality image extracted manually. Because the color and texture in the image of field plant is nonuniform, the artificial features in the image are poorly distinguishable and mismatching could occur as a result. This would compromise the accuracy of the depth of the map. While the supervised learning-based convolution neural network (CNN) is able to estimate the depth of each pixel in plant image directly, it is expensive to annotate the depth data. In this paper, we present a depth estimation model based on the self-supervised learning to phenotype tomato canopy. The tasks of the depth estimation method were to reconstruct the image. The dense disparity maps were estimated indirectly using the rectified stereo pair of images as the network input, from which a bilinear interpolation was used to sample the input images to reconstruct the warping images. We developed three channel wise group convolutional (CWGC) modules, including the dimension invariable convolution module, the down-sampling convolution module and the up-sampling convolution module, and used them to construct the convolutional auto-encoder - a key infrastructure in the depth estimation method. Considering the shortage of manual features for comparing image similarity, we used the loss in image convolutional feature similarity as one objective of the network training. A CWGC-based CNN classification network (CWGCNet) was developed to extract the low-level features automatically. In addition to the loss in image convolutional feature similarity, we also considered the whole training loss, which include the image appearance matching loss, disparity smoothness loss and left-right disparity consistency loss. A stereo pair of images of tomato was sampled using a binocular camera in a greenhouse. After epipolar rectification, the pair of images was constructed for training and testing of the depth estimation model. Using the Microsoft Cognitive Toolkit (CNTK), the CWGCNet and the depth estimation network of the tomato images were calculated using Python. Both training and testing experiments were conducted in a computer with a Tesla K40c GPU (graphics processing unit). The results showed that the shallow convolutional layer of the CWGCNet successfully extracted the low-level multiformity image features to calculate the loss in image convolutional feature similarity. The convolutional auto-encoder developed in this paper was able to significantly improve the disparity map estimated by the depth estimation model. The loss function in image convolutional feature similarity had a remarkable effect on accuracy of the image depth. The accuracy of the disparity map estimated by the model increased with the number of convolution modules for calculating the loss in convolutional feature similarity. When sampled within 9.0 m, the root means square error (RMSE) and the mean absolute error (MAE) of the corner distance estimated by the model were less than 2.5 cm and 1.8 cm, respectively, while when sampled within 3.0m, the associated errors were less than 0.7cm and 0.5cm, respectively. The coefficient of determination (R²) of the proposed model was 0.8081, and the test speed was 28 fps (frames per second). Compared with the existing models, the proposed model reduced the RMSE and MAE by 33.1% and 35.6% respectively, while increased calculation speed by 52.2%. [ABSTRACT FROM AUTHOR]

In order to sort out the development of deep learning methods in the field of human action recognition, this paper summarized the most representative models and algorithms in this field in recent years. Firstly, it described in detail the latest achievements, advantages and disadvantages and network structure of deep learning methods in video pre-processing stage based on the task flow of human action recognition. Then, it introduced two kinds of datasets related to human action recognition. Finally, it discussed and prospected the future research direction of human action recognition. [ABSTRACT FROM AUTHOR]

The purpose is to deal with the problems of excessive pruning of important convolutional layers caused by equalscale pruning, retain many redundant parameters and large loss of accuracy. This paper integrated sensitivity analysis on the basis of FPGM pruning strategy for network pruning. The algorithm used precision feedback to analyze the importance of each convolutional layer, control the single-layer pruning ratio, analyze the impact of different pruning ratios in each layer on accuracy loss, and obtained the sensitivity of each convolutional layer. It could be combined with FPGM to analyze the importance of the convolution kernel and cut the unimportant convolution kernel according to the pruning ratio of the sensitivity to complete the neural network pruning. The experimental results show that the accuracy of this method only drops by l. 56% and 0. 11 % when the pruning rate on MobileNet-v 1 and ResNet50 is 50%. This method has a higher pruning rate and lower calculation amount on ResNet50 under the same accuracy loss. [ABSTRACT FROM AUTHOR]

Aiming at the problems of complex background, variety of change types, missing detection and rough boundary recognition in high-resolution remote sensing image of the same region, this paper proposed a high-resolution remote sensing image building change detection algorithm based on Siam-UNet++network. The algorithm used UNet++as the backbone extraction network. In the encoder phase, it applied the Siam-diff structure to extract the change features of the two sequential images, and employed the attention mechanism after the up sampling and lateral jump path connection in the decoding stage to highlight the building change features and inhibit the network learning from other types of features. Meanwhile, it used the MSOF strategy to weight and fuse feature information of different semantic levels, which improved the accuracy of building change detection. Finally, it adopted a sliding window method to predict large-scale remote sensing images, reducing the hole pattern generated by the change result map during the splicing process. The experimental results demonstrate that proposed algorithm shows better performance than other models. [ABSTRACT FROM AUTHOR]

Network representation learning aims to transform the nodes in the network into a low-dimensional vector space while maintaining the inherent properties of the network. Most of the existing methods are aimed at normal networks, ignoring the particularity of attribute bipartite networks and the highly non-linear characteristics of the network. To solve the above problem, this paper proposed a deep attributed bipartite network embedding method incorporating with multilevel structure information. Specifically, the algorithm introduced an extended weight matrix to fuse the explicit and implicit structure of the bipartite network with attribute information. Then, it utilized a deep auto-encoder model to capture the highly non-linear characteristics of the network. To maintain the global network structure, a deep auto-encoder reconstructed the second-order proximity. Meanwhile, the deep auto-encoder used the first-order proximity of the nodes as supervisory information to maintain the local network structure. Finally, the algorithm performed joint optimization to get the final representation vectors of nodes. The model was executed on the four datasets of Yelp, Douban Book, Douban Movie and MovieLens. Compared with the latest benchmark method, the average values of F1@10, MAP@10,MRR@10 and NDGG@10 of this model have improved by 4.29%、5.63%、6.26%、4.21%. [ABSTRACT FROM AUTHOR]

In the scene of document verification，the standard deep learning face recognition method has low accuracy and poor real-time performance on embedded devices. To solve these problems，this paper proposes a modified efficient convolutional neural network（CNN）called Lightnet and adopts the transfer learning method. Lightnet is an efficient CNN module composed of depthwise separable convolution，linear bottleneck structure and attention module. After introducing the loss function AM-Softmax with additive angle margin， the network model can effectively solve the problems of redundancy parameter and vast calculation for standard CNN in the foundation of ensuring the high accuracy of face recognition. The transfer learning method can enhance the scene-identity face matching performance by freezing all the convolution layer weights of the pre-trained model and fine-tuning training in the self-made scene-identity face matching dataset. The experimental results show that the designed efficient scene-identity face matching algorithm has achieved good results in terms of verification accuracy，parameters and verification speed，and has good robustness in life scenarios. [ABSTRACT FROM AUTHOR]

In order to solve the problem that the recommendation model cannot mine the diversity of user interest and capture the order information between the sequence of user behavior, and the interaction occurs at the element level rather than between the feature vectors, etc., based on multiple attention mechanism and location information, this paper proposed the x DMAL FM. Firstly, it extracted the feature depth from different subspaces by using the multi-head attention mechanism, and then location information could capture the sequential relationship between user behavior sequences. Finally, it used three public datasets to conduct comparative experiments and the AUC index to evaluate the results. The experimental results show that the proposed algorithm has better recommendation performance than the xDeepFM model and show the effectiveness and feasibility of the proposed algorithm. [ABSTRACT FROM AUTHOR]

In recent years, as deep learning has achieved good results in many fields, deep learning has also begun to be applied to recommendation systems, such as NFM models and DeepFM models that use deep learning technology to capture high-level feature interactions. However, considering the changes in the external environment and internal perception, the user' s interest should also change dynamically over time, and the combination based on the original features may not necessarily learn effective feature interaction. This paper attempted to build a new model FG_DRFwFm, which could learn the interaction of low-level and high-level features of multiple feature domains and dealt with long-term changes in user interest. It constructed the training features by constructing new features based on the original features and splicing them together, which could better learn effective feature interaction. Finally, the proposed model compared the recommendation effect with multiple advanced CTR algorithms on the MovieLens data set. The experimental results show that the proposed model achieves better results. [ABSTRACT FROM AUTHOR]

Considering federated learning under mobile edge computing, where a global server connects a large number of mobile devices through the network to jointly train a deep neural network model. The data distribution shift caused by global class imbalance and device local class imbalance leads to the performance degradation of the standard federated averaging ( FedAvg) algorithm. This paper proposed a device selection algorithm based on the combinatorial multi-armed bandit ( CMAB) as an online learning algorithm framework, and designed a class estimation scheme to form a nonlinear reward function. Based on class estimation scheme and CMAB, in each round of communication, the global server selected the device subset with class imbalance that could be best complementary with test performance deviation across classes of global model from last communication round. Therefore, the current aggregated global model could achieve more balance and better test performance per class, and also faster convergence and more stable training dynamics. Extensive numerical results demonstrate the influence of different parameters on FedAvg under class imbalance, and verify the effectiveness of the proposed algorithm. [ABSTRACT FROM AUTHOR]

Aiming at the problem of high pedestrian missing rate in a crowded, this paper designed a new balance strategy. Secondly, it used metric learning method to improve the effect of pedestrian semantic extraction, and designed a new distance measurement method. Finally, combining with the extracted pedestrian semantic information, it designed a new non-maximum suppression algorithm. Compared with the current pedestrian detector, the proposed detector is better than other pedestrian detectors in Citypersons and Crowdhuman datasets. Experiments also prove the effectiveness of the metric learning method in pedestrian detection. [ABSTRACT FROM AUTHOR]

Visual odometry (VO ) estimates the pose of a mobile robot by analyzing the image flow captured by the equipped cameras. In order to analyze the development of VO algorithms, this paper reviewed the related technologies of VO and the up-to-date research state combined with some advanced VO systems. Firstly, this paper described the concept and the evolution of VO, and introduced the mathematical description and the classification of VO . Then, it analyzed the key technologies of VO in details, including feature selection, motion estimation and drift reduction. In addition, it also introduced the latest deep learning based VO. At last, it discussed the existing problems and prospected the development trend of VO. [ABSTRACT FROM AUTHOR]

With the advancement of remote sensing technology, the high spatial resolution remote sensing image contains rich special information with a great detail. At the same time, the complexity of high spatial resolution remote sensing images also requires higher the classification technology of remote sensing images. However, in the face of high spatial resolution remote sensing image more obvious geometrical structure and the more rich texture characteristics, how to design rational system of characteristics, select the appropriate sorting algorithms to accurately and quickly grasp the number of rural land of building and its distribution status, are of great significance to balance urban and rural areas, save land, and realize sustainable development. This will help in exploring the application of deep learning model in high spatial resolution remote sensing image building extraction, and have research significance for improving the classification accuracy of high resolution remote sensing image. In this paper, the semantic segmentation model (SegNet) was used for extracting buildings. SegNet is mainly composed of encoder network, decoder network and pixel-wise classification layer. The encoder network transforms high-dimensional vectors into low-dimensional vectors, enabling low-dimensional extraction of high-dimensional features. The decoder network maps low-resolution feature maps to high spatial resolution feature maps, realizing the reconstruction of low-dimensional vectors to high-dimensional vectors. The softmax classifier separately classifies each pixel, which outputs the probability that each pixel belongs to each class. In this paper, a 3000 pixel × 3000 pixel and two 2000 pixel × 2000 pixel slices were taken from the global remote sensing image of Bazhou City, Hebei Province as training samples, and a 3000 pixel × 3000 pixel slice was taken as the verification sample. In this paper, five comparative experiments were used to extract the buildings, including PSPNet, support vector machine, random forest, ISO clustering and maximum likelihood method. The confusion matrix of each classification method was obtained by calculating the difference between the classification results of the comparison experiment and the real value. From the traditional classification algorithm to the shallow learning algorithm to the deep learning algorithm, the Kappa coefficient and overall accuracy of classification kept constantly increasing, among which SegNet semantic segmentation algorithm based on the deep convolutional network performed better than the other five algorithms in extracting buildings from high spatial resolution remote sensing image. The Kappa coefficient and the overall accuracy of SegNet semantic segmentation algorithm were 0.90 and 96.61%, respectively, and the ground truth value was basically the same as the classification result. The F1Score of building extraction of SegNet semantic segmentation algorithm based on deep convolution network was 0.91, but the other five algorithms were below 0.87. SegNet had the lowest error rate of 9.71% for buildings, indicating that the ability to identify buildings of semantic segmentation algorithm from high spatial resolution remote sensing was superior to traditional classification algorithms, shallow layer learning algorithms based on machine learning, and PSPNet semantic segmentation algorithm based on deep convolution network. The Kappa coefficient and overall accuracy of the remaining five classification algorithms were respectively below 0.83 and 94.68%, and the difference between the ground truth value and the classification result was relatively large. SegNet can not only make use of spectral information but also make full use of abundant spatial information. During SegNet training, more essential features can be learned, and more ideal features suitable for pattern classification were finally formed, which can enhance the ability of convergence and generalization of the model and improve the classification accuracy. Traditional classification algorithms, such as ISO clustering and maximum likelihood method, failed to make use of the rich spatial information of the high-resolution remote sensing image, so the accuracy was relatively low. Due to limited computing units and large amount of high spatial resolution remote sensing image data, shallow layer learning algorithms based on machine learning such as support vector machines and random forest cannot effectively express complex features of ground objects, so their advantages are not obvious in building extraction from the high spatial resolution remote sensing images. The experimental results showed that the SegNet based on deep learning has the best performance, and it has important theoretical significance to explore the application of deep learning model to remote sensing image classification methods. At the same time, the research results also provide a reference for improving the classification accuracy of high resolution remote sensing images. [ABSTRACT FROM AUTHOR]

In order to solve drawbacks of the missing preference from the user and low accuracy in personalized aesthetic ranking methods, this paper proposed a SVMRank based model to automatically rank the input images. The framework took as input a series of specific photos that users prefer, then deployed DCNN to extract the deep features and compared them with ones that were extracted from dataset to establish an user-specific aesthetic training dataset. Later, it used SVMRank to learn a customized hyper plane to produce a user-specific personalized aesthetic ranking. This paper conducted two experiments at mean time. It invited several users to evaluate the performance of user-specific aesthetic ranking, and tested the accuracy on binary-classification.The experiment results show that the framework performs well on predicting user’s preference as well as on classifying the images from high to low quality. In conclusion, the proposed algorithm is an effective personalized ranking algorithm. [ABSTRACT FROM AUTHOR]

In traditional cover selection image steganography, the relationship between image and secret information needs to be constructed manually, and the robustness is not strong. In response to this problem, this paper proposed a cover selection image steganography algorithm. This algorithm combined the idea of image retrieval, used the discriminator of StarGAN as the feature extractor, mapped the extracted features into secret messages, and directly constructed the relationship between image and secret messages. In order to automatically update the mapping relationship database, this solution used feature update, searching image update and other methods to update the relationship database. Experiments show that this scheme has a higher steganography capacity and better robustness. When the image quality is not high, the extraction accuracy is about 6% higher than the existing methods. [ABSTRACT FROM AUTHOR]