Search

According to the problem of the current RGB-D feature extraction network used for pose estimation was too large, this paper proposed a lightweight improved XYZNet RGB-D feature extraction network. First, this paper designed a lightweight sub-network BaseNet to replace ResNet18 in XYZNet, which made the network scale significantly reduced and obtained more powerful performance. Then, this paper proposed a re-parameterized multiscale convolutional attention (Rep-MSCA) sub-module based on depth separable convolution, which enhanced the ability of BaseNet to extract contextual information of different scales, and constrained the amount of parameters in the model. Finally, in order to improve the geometric feature extraction ability of PointNet in XYZNet with a small parameter cost, this paper designed a re-parameterized residual multi-layer perceptron (Rep-ResP) module. The floating point operations (FLOPs) and parameters of the improved network are 60.8% and 64.8% lower, the inference speed is 21.2% higher, and the accuracy of the mainstream datasets LineMOD and YCB-Video is 0.5% and 0.6% higher. The proposed model is more suitable for deployment in scenarios where hardware resources are tight. [ABSTRACT FROM AUTHOR]

BACKGROUND: Traditional 3D dental segmentation methods usually utilize predefined spatial geometric features, such as curvature and normal vectors, as the reference information for tooth segmentation. OBJECTIVE: To propose an algorithm for complex 3D dental segmentation and deeply explore the correlation between segmentation results and application scenarios. METHODS: A 3D dental segmentation algorithm based on dual stream extraction of structural features and spatial features was established, and the modular design of split flow was used to avoid feature confusion. Among them, the attention mechanism on the structural feature flow was used to capture the finegrained semantic information required for tooth segmentation, and the Tran Net based on the spatial feature flow was used to ensure the robustness of the model to complex tooth and jaw segmentation. This algorithm verified its effectiveness and reliability based on clinical datasets including healthy dental jaws and complex dental jaws such as missing teeth, malocclusion and dentition crowding. The segmentation performance of the model was measured in terms of overall accuracy, mean intersection over union, and directional cut discrepancy. RESULTS AND CONCLUSION: The overall segmentation accuracy of this algorithm in the clinical data set is 97.08%, and the segmentation effect is superior to that of other competitive methods from the qualitative and quantitative perspectives. It is verified that the structural feature flow designed in this paper can extract more precise local details of tooth shape from coordinate and normal information by constructing an attention aggregation mechanism, and the spatial feature flow designed in this paper can ensure the robustness of the model to complex teeth such as missing teeth, dislocated teeth, and crowded dentition by constructing a transformation network (Tran Net). Therefore, this tooth segmentation algorithm is highly reliable for clinicians' practical reference. [ABSTRACT FROM AUTHOR]

Effectively representing features of modalities is a hot issue in cross-modal recipe retrieval. The current methods generally adopt two independent neural networks to extract the features of images and recipes respectively, achieving retrieval through cross-modal alignment. However, these methods mainly focus on the intra-modal information and ignore the intermodal interactions, resulting in the loss of some effective modality information. To address the problem, this paper proposed a cross-modal recipe retrieval method to enhance modality semantics through multimodal encoders. Firstly, it used a pre-trained model to extract initial semantic features of images and recipes and utilizing modality alignment to reduce the inter-model differences. Secondly, it employed the pairwise cross-modal attention to repeatedly reinforce the features of one modality by using features from another modality, extracted valid information. Thirdly, it used the self-attention mechanism to modal the internal features of modalities to capture rich modality-specific semantic information and potential associations. Finally, it introduced the triplet loss to minimize the distance between similar samples, achieved cross-modal retrieval learning. Experimental results on Recipe 1M dataset show that the proposed approach outperforms the current mainstream methods in terms of median ranking(MedR) and recall rate at top K(R@K), providing a powerful solution for cross-modal retrieval tasks. [ABSTRACT FROM AUTHOR]

Autonomous vehicle can calculate the movement track of surrounding vehicles according to the track prediction algorithm, and make response to reduce driving risk, while the traditional track prediction model will produce large errors in the case of long-term series prediction. To address this issue, this paper proposed a trajectory prediction model based on the Informer algorithm, and used the publicly available dataset NGSIM to conduct experimental analysison. Firstly, it filtered the original data by using symmetric exponential moving average method(sEMA), and added a joint normalization layer to the original Informer encoder to extract features from different vehicles, reducing the motion error between different vehicles, and improving the prediction accuracy by considering the speed information of the vehicle itself and the vehicle movement information of the surrounding environment. Finally, it got the vehicle trajectory distribution at the future time through the decoder. The results show that the trajectory prediction error of the model is less than 0.5 m. Through the analysis of MAE and MSE results of trajectory prediction, when the prediction time exceeds 0.3 s, the trajectory prediction effect of Informer model is obviously better than other algorithms, which verifies the effectiveness of the model and algorithm. [ABSTRACT FROM AUTHOR]

With the rapid development of big data and Internet of Things (IoT) technologies, the application scope of multidimensional time series data has become increasingly widespread. Faced with a large amount of complex time series data characterized by non-linearity and high-dimensional redundant features, traditional time series analysis methods struggle to effectively address the complexity of multi-dimensional time series with high-dimensional features, resulting in suboptimal predictive performance. To address these issues, this paper proposes the GGInformer model, which improves upon the Generic Algorithm and Informer model while incorporating the GRU network. This model not only efficiently extracts key features from multidimensional time series but also effectively addresses longterm dependency issues. To validate the predictive capability of the model, experiments are conducted on two real datasets and three pubic benchmark datasets, all of which demonstrated superior performance compared to the baseline models. Specially, compared to the Informer baseline model, the GGInformer model achieves reductions in Mean Squared Error (MSE ) values of 22 %, 13 %, 20%, 23 %, and 38 % across the five datasets. The experimental results indicate that the GGInformer model can effectively address the complex feature extraction challenges of multidimensional time series data and further enhance time series prediction capabilities. [ABSTRACT FROM AUTHOR]

Multimodal emotion recognition in conversations（MERC） is a task to predict the emotional category of the discourse in a dialogue based on its textual, audio, and visual modality. Existing studies focus on multimodal feature extraction and fusion of discourse context without fully considering the utilization of emotional features of different speakers. Therefore, this paper proposed a model of multimodal dialogue emotion recognition based on a consistent graph convolutional network. The model first constructed a graph convolutional network of multimodal feature learning and fusion, and obtained the context features of each discourse. On this basis, the average features of the speaker in the complete dialogue as the constraint to make the model learn more reasonable discourse features, so as to improve the performance of predicting emotion class. The paper compared with other baseline models on two benchmark datasets IEMOCAP and MELD. And the results show that the proposed model is superior to the other models. In addition, the paper verifies the consistency constraints and other components of the model through ablation experiments. [ABSTRACT FROM AUTHOR]

In order to solve the problem that class-imbalance data of clinical medical tables tend to have an impact on the model and that the performance of deep learning framework is difficult to match that of traditional machine learning methods when processing scale data tasks, this paper proposed a layer by layer Transformer (LLT) network model based on cascaded under-sampling. LLT deleted the most types of data layer by layer by cascade under-sampling method to achieve the balance of data categories and reduced the impact of class-imbalance data on the classifier. Moreover, LLT used attention mechanism to carry out correlation evaluation on the features of the input data to achieve feature selection, refined the feature extraction abi-lity and improved the model performance. This paper used RA (rheumatoid arthritis) data as test samples. Experimental results show that, on the premise of not changing the sample distribution, the recognition rate of a few categories is increased by 6.1% by the proposed cascade under-sampling method, which is 1.4% and 10.4% higher than that of the commonly used NEARMISS and ADASYN respectively. The accuracy of the RA tabular data and the F1-score index of LLT reach 72.6% and 71.5%, the AUC value is 0.89, the mAP value is 0.79, and the performance exceeds the current mainstream tabular data classification models such as RF, XGBoost and GBDT. This paper also visualized the model process and analyzed the characteristics affecting RA. It has a good guiding significance for the clinical diagnosis of RA. [ABSTRACT FROM AUTHOR]

The semi-supervised deep neural network modeling method has been widely applied in soft sensor, but the network based on hierarchical training only excavates the effective information of each input layer in the feature extraction process, ignoring the loss of effective information of the original input and accumulating it layer by layer, resulting in low accuracy of feature representation of the original input. In addition, the lack of spatiotemporal information related to the mining process can also lead to model performance degradation. This paper proposed a semi-supervised dynamics deep fusion neural network (SS- DDFNN) method. This method reconstructed the original input data and predicted quality variables at each layer of the feature extraction network. By using the reconstructed original input error in pre-training loss, it reduced the loss of effective information from the original input. Simultaneously it incorporated attention mechanism and t-distribution random neighborhood embedding to extract spatiotemporal related information, and established a gated neural network quality prediction model using extracted features. The experimental results show that compared to the SAE, GSTAE, and SIAE models, the proposed method has improved prediction accuracy by 2.8%, 1.1%, and 0.9% in the case of a debutanizer, respectively. In the industrial polyethylene production case, it has increased by 2.7%, 1.0%, and 0.7% respectively. The experimental results show that the pro- posed method is effective. [ABSTRACT FROM AUTHOR]

Most of the existing user profile methods lack different granularity text information representation, and there is a noise problem in the feature extraction stage, resulting in the inaccurate construction of the profile. To address these issues, this paper proposed a user profile method based on multi-granularity information fusion, called UP-MGIF. Firstly, it integrated the character-level granularity and the word-level granularity representation vectors in the embedding layer to expand feature content. Secondly, based on the improved bi-directional gated recurrent unit network(Bi-GRU), it designed a hybrid feature extraction model called Bi-GRU-DAE-Attention by combining denoising autoencoder (DAE) and attention mechanism to achieve feature denoising and semantic enhancement. Finally, it input the robust feature vectors into the classifier to achieve user profile generation. Experiments show that the user profile generation method achieves higher classification accuracy than other baseline methods on two profile datasets in the medical and Internet domains, and validate the effectiveness of each module through ablation experiments. [ABSTRACT FROM AUTHOR]

Aiming at the issue of insufficient utilization of temporal information, speaker information, and multi-modal information in existing conversational emotion recognition methods, this paper proposed a multi-modal temporal-aware DAG model(MTDAG). The designed temporal-aware unit optimized the discourse weight setting in chronological order and collected historical emotional cues to achieve more effective utilization of temporal and historical information based on recency effect. The context and speaker information fusion module achieved the full utilization of discourse information by extracting the deep joint information of contextual context and speaker self-context. By setting the DAG subgraphs to capture multi-modal information and constrain the interaction direction, the model achieved full utilization of multi-modal information while reducing the introduction of noise. Extensive experiments conducted on two benchmark datasets, IEMOCAP and MELD, demonstrate that the model exhibits excellent performance in emotion recognition. [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]

In order to solve the problem that the current three-dimensional segmentation methods ignored the effective use of geometric features when mining point cloud features, this paper proposed the point cloud instance segmentation network 3D-BSNet with bilateral features and similarity measure. The network mainly consisted of two parts, such as bilateral feature learning and lightweight similarity measurement. Firstly, this paper proposed the bilateral feature extraction module with 3D-UNet based on submanifold sparse convolution and multi-layer perceptron. This module could extract semantic and geometric features of point cloud data after voxelization. Secondly, this paper designed the bilateral attention mechanism combining channel dimension and spatial dimension. This module could reduce the information loss in the process of bilateral feature aggregation. Finally, this paper developed the lightweight similarity measurement module. This module could obtain the similarity between neighboring point clouds in the high-dimensional embedded feature space and even generate the segmentation results of fine-grained instances. Experiments show that 3D-BSNet performs well on S3DIS and Scannet（v2） dataset, and the average accuracy rate on Scannet (v2) is 3.3% higher than that of SSTNet, which effectively improves the precision of 3D instance segmentation. [ABSTRACT FROM AUTHOR]

Most of the Encryption-then-Compression (ETC) methods can only obtain several limited fixed compression ratios. However, arbitrary encryption compression ratios instead of limited fixed compression ratios are more suitable for practical requirements. To this end, this paper proposed a lossy compression algorithm for encrypted color images with arbitrary compression ratios. It combines the uniform and random downsampling to compress the encrypted images, obtaining arbitrary compression ratios of an encrypted image. The receiver receives the compressed sequence of the encrypted image and obtains the decrypted image by decompression and decryption. The proposed scheme then characterizes the lossy reconstruction of the original image from the decrypted image as an optimization problem with the downsampling compression-based constraint. This scheme designs a convolutional neural network-based image reconstruction model for lossy ETC to resolve this problem, which is denoted the ETC-oriented Reconstruction Network (ETRN). ETRN consists of shallow feature extraction (SFE), residual in residual (RIR), residual content supplementation (RCS), and down-sampling constraint (DC). The experimental simulation results show that the encrypted color image lossy compression algorithm proposed in this paper can obtain excellent compression and reconstruction performance, which fully demonstrates the feasibility and effectiveness of this method. [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]

The level of face recognition technology continues to improve, and has achieved an ideal recognition rate in identity authentication, human-computer interaction and other industrial applications, and the market scale keeps growing. However, the occlusion problem in real scenes has not been completely solved. How to restrain or eliminate the negative impact of occlusion on key facial feature, is one of the hot spots in the field of face recognition. Aiming at the problem of lack of face structure information caused by occlusion, this paper reviews the occluded face recognition datasets and the occluded face recognition methods. Firstly, this paper introduces and analyzes some important new occluded face recognition datasets; Secondly, this paper summarizes the traditional learning methods and deep learning based methods to solve occlusion problems, and emphasizes the deep learning based robustness feature extraction and occluded facial information recovery. Finally, this paper summarizes and analyzes the advantages and disadvantages of relevant methods, points out the problems and challenges of occluded face recognition, and prospects the future research directions. [ABSTRACT FROM AUTHOR]

[Objectives]Aiming at the difficulty of identifying pests and diseases in the real environment of bergamot in the field, this paper proposed a method for detecting pests and diseases of bergamot based on improved SSD(single shot MultiBox detector)algorithm—SSD-Res50-3C algorithm. [Methods]The backbone network partially replaced the original VGG16 network with the ResNet50 network, and enhanced the model's ability to extract the characteristics of pests and diseases of bergamot in the real background of the field. A lightweight and efficient feature fusion module was added before predicting the feature layer to improve the multi-scale feature fusion capability of the SSD algorithm, and further improved the anti-interference ability of the SSD algorithm in the real background of the field. [Results]The mean average precision of SSD-Res50-3C algorithm was 92.86%, which was 6.61% higher than that of the original SSD algorithm, and the FPS(frames per second)reached 64.1. Compared with YOLO v3, YOLO v4, YOLO v5x6, Faster R-CNN, and EfficientDet-D3 models, the mean average precision of SSD-Res50-3C algorithm was 6.41%, 2.01%, 0.79%, 0.58%, and 5.10%, respectively, and FPS was 16.20, 40.28, 24.40, 36.20, and 54.84, respectively. [Conclusions]The pests and diseases detection method of bergamot based on the improved SSD algorithm proposed in this paper could weaken the interference information of the real environment in the field, accurately identify the target of pests and diseases of bergamot in the real environment of the field, and provide a new idea for the detection of pests and diseases of bergamot in the real environment of the field. [ABSTRACT FROM AUTHOR]

Autonomous navigation is an important guarantee for intelligent agricultural machinery to complete harvesting operations. Navigation line recognition is an important step in achieving autonomous navigation of intelligent agricultural machinery. Real-time and accurate detection of harvesting navigation lines can effectively accelerate the progress of operations and reduce crop loss rates. Visual sensors have the advantages of low cost and rich information acquisition, making them widely used in route detection. Frequent changes in lighting under cloudy weather make it difficult for a single feature to cope with wheat field environments. In this study, a wheat harvesting boundary detection method based on multi texture feature fusion was proposed. In the harvested area, the characteristics of low stubble and no top wheat ears result in differences in uniformity, density, and other aspects when harvesting under both light and backlight compared to the non-harvested area. At the same time, due to the exposed wheat stubble in the harvested area, it appears more regular on the image, and its gradient direction is more consistent than that in the non-harvested area. Therefore, a two-dimensional feature vector composed of image entropy and directional gradient was constructed to classify the harvested and unharvested areas of the wheat field, and then the harvesting boundary was extracted. Due to the susceptibility of image entropy features to perspective phenomena, this paper compared the entropy features of different regions in the image when fixing the camera installation angle and height, and ultimately selected the middle region of the image as the region of interest. After analyzing the characteristics of image entropy feature extraction, a histogram statistical method based on sliding windows was proposed to accelerate the speed of image entropy feature extraction. By further dividing the window to be calculated into several sub windows, sliding and combining these sub windows significantly reduced the computation. The entropy feature extraction in this paper took 0.53 seconds, which is 49.52% faster than the traditional method of directly extracting entropy from the entire window. After extracting twodimensional features, based on the distribution characteristics of the feature histogram and combined with the maximum entropy threshold segmentation algorithm, the wheat field image was preliminarily classified. For entropy features, low noise points affected the threshold segmentation effect. By comparing the impact of discarding different proportions of low entropy points on the segmentation threshold, this paper discarded 10% of low entropy points, enhancing the threshold segmentation effect. The method of removing small connected regions was used to eliminate the misclassified regions in the initial classification binary image. Then, the Canny operator was used to extract edge contour points distributed near the harvesting boundary. Due to the consistent direction between the harvesting boundary and the agricultural machinery's forward direction, the Ransac algorithm was used to restrict the fitting line in the region, obtaining the accurate harvesting boundary. In order to verify the feasibility of the method proposed in this article, wheat field images under different lighting conditions were collected, including 2 200 weak light images, 758 local strong light images, and 1 134 strong light images. After selecting the region of interest, the size of the image to be processed was 240 pixels×1 280 pixels. The algorithm in this paper took an average of 0.88 s to process each image on laptops. The detection rates under weak light, local strong light, and strong light were 90.41%, 88.26%, and 89.68%, respectively, with an average of 89.45%. Compared with the traditional method of using Adaboost algorithm for machine learning, the detection speed of this algorithm was improved by 73.89%, and the detection rate was increased by 46.19 percentage points, 46.00 percentage points, and 49.64 percentage points under weak light, local strong light, and strong light, respectively. Compared with traditional algorithms, the method proposed in this paper significantly improved real-time performance and detection rate, which can provide reference for the detection of field agricultural machinery navigation lines. [ABSTRACT FROM AUTHOR]

In order to solve the problems of feature extraction and temporal translation in sign language translation methods, this paper proposed an AGCN-T sign language translation network that combined AGCN and Transformer temporal model. The AGCN module learnt the interactive spatial dependency information of skeleton nodes in sign language actions. The Transformer temporal module captured the time relationship feature information of the sign language action sequence and translated it into understandable sign language semantic information. In addition, in the preprocessing part, this paper proposed a key frame extraction algorithm for moving windows, and used MediaPipe pose estimation algorithm to extract the skeleton of key frame image sequences. The experimental results show that the proposed method achieves a word error rate of 3. 75 % and an accuracy of 97. 87% in the large-scale Chinese continuous sign language dataset CCSL, which is superior to other advanced sign language translation methods. [ABSTRACT FROM AUTHOR]

In the past object detection tasks,many studies construct feature optimization algorithms through spatial and channel information. How to utilize the information distribution of the largest and average features in the channel and space for feature extraction is still a challenge. In response to the above challenges, this paper constructed a dual channel space interdependent network to extract information-dependent features in space and channels. The key was to obtain the vital information distribution on the maximum feature and the average feature. In order to optimize the detection accuracy and computational efficiency of the network, this paper constructed an asymmetric weight-sharing convolution network, using the same asymmetric convolution kernel for joint training to realize weight sharing, which greatly enhanced the robustness of the convolutional kernel parameters. This paper conducted experiments on PASCAL VOC and MS COCO datasets using YOLOv4, YOLOv5, and Efficient Det as detection baselines to verify the effectiveness of the dual channel space interdependent network and the asymmetric weight-sharing convolution network. The highest accuracy increases by l. 98% and 2. 6% on the two datasets. [ABSTRACT FROM AUTHOR]

The layout of document images is complex and distribution of object sizes is uneven, currently, most of detection methods ignore multimodal information and global dependencies. Therefore, this paper proposed a multi modal document object detection method based on vision and text. Firstly, this method explored the fusion strategy of multimodal features. In order to utilize textual features, it converted text sequence information of the image into two-dimensional representation. After the initial fusion of text features and visual features, it input the fused features to backbone network to extract multiscale features, and repeatedly integrated textual features during the extraction process, so as to realize deep fusion of multimodal features. Next, to ensure the detection accuracy of small and large objects, this paper designed a pyramid network. The lateral connection could concatenate feature maps of the same spatial size from the bottom-up pathway and the top-down pathway in channel, so as to achieve the propagation between high-level semantic information and low-level feature information. The experimental results on large public dataset PubLayNet show that the detection accuracy of this method reaches 95. 86%,and it has a higher accuracy than other methods. This method not only realizes the deep fusion of multi modal features, but also enriches the fused multimodal feature information, and it has good detection performance. [ABSTRACT FROM AUTHOR]

Local feature information plays an important role in image segmentation, however, the referring image segmentation task is dependent on the text expression, so it is impossible to extract local feature information directly from the original reference image. In order to solve this problem, this paper proposes a specific noun-guided local feature extraction deep neural network model（NgLFNet). The NgLFNet model can automatically mine the local feature information of the object to be segmented according to the key nouns in the input text expression. Specifically, the model first obtains key nouns in text through sentence analysis; secondly, extracting corresponding features through text and image encoders, and using the key nouns to obtain local features of high-interest regions through the multihead attention mechanism; then the multi-modal features are gradually fused to learn; Finally, the decoding and correction module uses the obtained local features to perform more detailed corrections on the prediction mask to obtain the final result. The method in this paper is compared with a variety of mainstream referring segmentation methods and the experimental results show that the proposed method improves the accuracy of text-based instance segmentation task. [ABSTRACT FROM AUTHOR]

Object surface defect detection technology is an important subject in the field of industrial quality inspection, which is of great significance to industrial production. This paper mainly summarized the surface defect detection techniques based on machine vision in recent years. Firstly, it listed several application scenarios of defect detection in the industry. Secondly, from the perspective of feature extraction and classification algorithms, it expounded on traditional machine vision methods. Then, this paper focused on the classical neural network structure and the latest development of defect detection algorithms and introduced two commonly used optimization methods of the defect detection algorithm. Finally, it analyzed three major challenges in the field of defect detection; real-time problems, small sample problems and small target problems. The purpose was to provide useful references and context for the research of industrial surface defect detection. [ABSTRACT FROM AUTHOR]

Aiming at the problems of local structure disconnection and inaccurate detail restoration in existing image restoration algorithms, this paper proposed an inpainting algorithm based on semantic prior and dual channel feature extraction（SPDCE）. The algorithm used semantic prior network to learn the semantic information and context knowledge of the missing area, predicted the missing area, and enhanced the local consistency of the generated image;then, the dual channel feature extraction network fully mined image information, enhanced the perception and utilization of texture details; next, the context feature adjustment module could capture and encode rich semantic features at multiple scales, thereby generating more realistic image views and finer texture details. After conducting experimental verification on the datasets CelebA-HQ and Places2, the results showed that compared with commonly used algorithms, the SPDCE algorithm improved the peak signal ratio（PSNR） and structural similarity（SSIM） by 1.6dB to 1.73dB and 3.1% to 9.9%, while L1 loss decreased by 15.2% to 27.8%. Experimental results show that the repaired image of the proposed algorithm has a more reasonable structure and richer details, and the inpainting effect is better. [ABSTRACT FROM AUTHOR]

The process of drum drying silk has complex characteristics of multivariability, strong coupling, and nonlinearity. Traditional Principal Component Analysis (PCA) method lacks strong nonlinear ability, and global modeling method is difficult to achieve accurate fault detection of the process. Therefore, this paper proposed a novel multi-block approach by integrating Autoencoder (AE) for feature extraction and PCA model. Initially, in order to capture local features, the process variables were divided into blocks according to the drying process principle of tobacco leaf. Secondly, autoencoder was used to extract the nonlinear features of each sub-block. Then, the corresponding PCA models are established for each sub-block, respectively. Lastly, the monitoring results of multiple subspaces were fused for decision-making by Bayesian inference. Two actual leaf silk drying cases were used for verification, and the results showed that the alarm rates of this method were as high as 91.67% and 98.21%. Compared to traditional PCA and AE-PCA detection methods, this algorithm could accurately reveal and characterize the overall operating status and local feature information of the drying process, improve the accuracy of anomaly detection in the drum leaf silk drying production process and achieve accurate alarm for quality anomalies, to ensure stable production of the drum leaf silk drying process. [ABSTRACT FROM AUTHOR]

To address the issue of large image super-resolution models with excessive parameters that are difficult to deploy, as well as the poor performance of existing lightweight image super-resolution models, this paper proposed an image super-resolution model based on adaptive attention fusion feature extraction network (AAFFEN). The model consisted primarily of a large kernel attention block and multiple efficient attention fusion feature extraction blocks. Firstly, the model extracted the shallow feature information using the large kernel attention block, and then a cascaded series of efficient attention fusion feature extraction block performed deep feature extraction, enhancement, refinement, and redistribution of the aggregated operations on the extracted shallow feature information. The efficient attention fusion feature extraction block consisted of three parts, such as progressive residual feature extraction module, channel contrast-aware attention module, and channel-spatial joint attention module. The proposed network could achieve better image super-resolution performance with fewer parameters, making it an excellent lightweight image super-resolution model. By evaluating the proposed method on popular benchmark datasets and comparing it with existing methods, the results show that the proposed method has more superior performance. [ABSTRACT FROM AUTHOR]

To address the problem of unbalanced computational resource consumption and accuracy of semantic segmentation models using attention mechanism, this paper proposed a lightweight attention enhancement algorithm for semantic segmentation. Firstly, it designed a striped dimensional attention mechanism based on the shape characteristics of objects in driving scenes, used striped pooling instead of traditional square convolution, and combined dimensionality reduction operations to extract long-range semantic associations in each dimension to cut down the model computation. Then it fused the attention on channel domain and spatial domain to form a lightweight multidimensional attention fusion module that could be superimposed and disassembled to extract feature information in all directions and further improve the model accuracy. Finally, it inserted the module into the ResNet-101 backbone based encoding-decoding network to guide the semantic fusion of high and low layers, correct the feature map edge information, and supplement the prediction details. The experiments show that the module has strong robustness and generalization ability, cutting about 90% of the number of parameters and 80% of the computation compared with the same type of attention mechanism, and the segmentation accuracy still achieves a stable improvement. [ABSTRACT FROM AUTHOR]

To solve the problem of inaccurate description and unreliable relation inference in group activity recognition, this paper focused on constructing a scene relationship graph for three aspects: individual, group, and scene, and proposed a scene relationship graph network（SRGN） for group activity recognition. This method included a feature extraction module, a scene relation graph inference module, and a classification module. The feature extraction module extracted individual features, group features, and scene features by convolutional neural network. To fully explore the impact of scene on individual and group descriptions, the scene relation graph inference module learnt individual features and group features by building individual-scene and group-scene relationship graphs in a two-branch framework. Scene graph inference took into account the influence of individual on group and introduced a cross-branch module. It used the classification module to classify individual features and group features for prediction. The experimental results show that the group recognition accuracy of the proposed method on volleyball and collective activity data sets is improved by 1. 1% and 0. 5%,respectively. It verifies the validity of the scene graph in describing individual feature and group feature. [ABSTRACT FROM AUTHOR]

In the context of the rapid development of network information, the demand of different music lovers for music information retrieval is also increasing, and music classification has become an important research subject. This paper proposes a fast classification method of DUPSO-DSVM folk songs, which combines dissipative uniform particle swarm optimization (DUPSO) with distance sorted SVM (DSVM). This method uses DUPSO algorithm to optimize the penalty coefficient C and kernel function parameter g of SVM, and uses DSVM algorithm to optimize the parameter optimization time of DUPSO algorithm. The experimental results show that, DUPSO-SVM algorithm has a classification accuracy of 84%. After using DUPSO-DSVM algorithm, the training time of the algorithm only accounts for 26.26% of the unused DUPSO-DSVM algorithm, but it still maintains a high classification accuracy. [ABSTRACT FROM AUTHOR]

Aiming at the problem of the low accuracy of fine retinal vessel segmentation, this paper proposed improved U-Net methods combining scalable cascade modules, Transformer and self-calibrated attention modules to improve the accuracy of fine retinal vessel segmentation. Firstly, this paper used scalable cascaded modules in the encoding stage to enable the learning of complex and variable retinal vessel structures. Secondly, in the decoding stage, this paper proposed a self-calibration attention mechanism, which used the multi-scale compression excitation module to adaptively recalibrate the blood vessel from channel and spatial of features, it could enhance the feature response of the target area, and suppressed the background noise. Finally, the Transformer feature extraction block improved the capability of feature space mapping. This proposed method tested on public datasets, i. e. the DRIVE and CHASEDB1. The experimental results of proposed method show that the accuracy of the retinal vessel segmentation from the two datasets reach 96. 49%/96. 67%, the sensitivity reach 83. 75%/83. 30%, the specificity reach 98. 28%/98. 01% and the AUC reach 0. 987 1/0. 987 2, respectively. The performance of proposed method is better than most of existing methods, and each modules can improve the ability of fine retinal vessel segmentation. [ABSTRACT FROM AUTHOR]

In the intelligent transformation of the mechanical manufacturing, a large amount of spare parts models is inevitably generated, which brings huge challenges to the efficient retrieval of data. Considering the user-friendly and lightweight characteristics of design sketches, this paper conducted model retrieval of mechanical parts based on design sketches by constructing a deep cross-domain representation model. Aiming at the problem of cross-modal information association between sketches and 3D models, this paper proposed a feature joint learning method, which controlled the differences within and between classes of retrieved models. By this method, feature descriptors could acquire single-domain features as well as fuse the cross-domain information, and thus, the method established a consistent association representation of cross-modal data in a co-embedding space. Finally, it constructed the index table by using hash coding to realize fast retrieval of massive data. The experimental results on the parts data show that the proposed retrieval method based on the design sketch can not only achieve the most accurate retrieval results, but also have higher retrieval efficiency among the comparison methods. The method not only improves the accuracy of cross-modal spare parts information retrieval, but also improves the efficiency of data management, thereby indirectly improving the efficiency and convenience of product design. The related systems have been applied in some enterprises and have received good feedback. [ABSTRACT FROM AUTHOR]

Dyskinesia of Parkinson' s disease (PD) can affect oral, pharynx, palatine muscles, and facial muscles, causing vocal cord tremor and facial dyskinesia, which makes it possible to identify PD patients by using voiceprints and facial features. To effectively utilize the above two features to improve PD detection rate, this paper proposed a multimodal circulant fusion model based on multiscale features and dynamic attention mechanism. Firstly, this paper designed a multi-scale feature extraction network to fuse semantic information of high and low level features to obtain complete feature information. Secondly, considering the correlation and complementarily between modalities in the process of multi-scale feature fusion, it proposed a dynamic attention mechanism algorithm to generate attention feature maps with different modal information as auxiliary conditions to each other, which could reduce the redundancy of information during feature fusion. Finally, multimodal circulant fusion model could obtain more efficient fusion features by calculating the Hadamard product between each row vector of circulant matrix and feature vector. The results of several sets of experiments on the self-built dataset show that the detection accuracy of the proposed method is as high as 96. 24%, better than state-of-the-art unimodal and multimodal recognition algorithms. The method can effectively distinguish PD patients from healthy people and lay the foundation for efficient identification of PD patients. [ABSTRACT FROM AUTHOR]

Aiming at the problem that the current visual SLAM method has poor robustness and high time consumption, which makes the system localization not accurate enough, this paper proposed a visual-inertial SLAM algorithm based on point-line feature fusion. Firstly, it improved the LSD extraction quality through short-line culling and approximate line-segment merging strategies to improve the speed and accuracy of line feature detection. Then, it integrated the point, line, and !MU data in the back-end optimization effectively and established the minimized objective function for optimization to obtain a more accurate camera pose. Finally, this paper conducted experimental verification on the EuRoC dataset and real corridor scenes. The experiments show that the proposed algorithm can effectively improve the quality and speed of line feature extraction while effectively improving the localization accuracy of the SLAM system and obtaining a richer point-line structure map. [ABSTRACT FROM AUTHOR]

In order to solve the problem that the siamese neural network is sensitive to position, complex background and intra-class differences due to the use of image-level feature metrics, this paper proposed a Dual Metric Siamese Neural Network (DM-SiameseNet) . DM-SiameseNet used image-level features and local features (local descriptors) to jointly represent each image, then learned feature maps based on two different levels of similarity measures, and finally used an adaptive fusion strategy to adaptively integrate two different measurement result represented by the level feature. Experimental results showed that the accuracy of the improved model was increased by 5.04% and 9.66%respectively, and was higher than the measurement methods that only use image-level feature representation or local descriptor representation on miniImageNet, TieredImageNet, Stanford Dogs, Stanford Cars and CUB-200 datasets. The experimental results prove that the proposed model not only considers the global features of the image, but also captures more effective local feature information in the image, which improves the generalization ability of the model. [ABSTRACT FROM AUTHOR]

There are many fields choose to fuse LiDAR point cloud and camera images for applying. The accurate extrinsic calibration is a prerequisite for fusing their information. The point cloud feature extraction is a key step in extrinsic calibration. However, the low resolution and low quality of point clouds can affect the accuracy of calibration results. To address these problems, this paper proposed a LiDAR and camera extrinsic calibration method based on edge correlation point cloud. Firstly, it extracted the edge correlation point cloud of the calibration plate by using dual echoes. Then, it obtained the corner points from the edge correlation point cloud by the optimization method, where the corner points were compatible with the actual calibrated plate size. Finally, it matched the corner points in the point cloud with corner points in the image, and solved the extrinsic parameter between the LiDAR and the camera by the multi-point perspective method. The experimental results show that the reprojection error of the method is 1. 602 px, which is lower than that of comparison methods, verifying the effectiveness and accuracy of the method. [ABSTRACT FROM AUTHOR]

To solve the problem of combustion state recognition in solid oxide fuel cell (SOFC) combustor, this paper proposed a combustion state recognition method based on attention mechanism and image feature pyramid. The method adopted the adaptive gamma correction with weighting distribution (AGCWD) to standardize the input images. It combined two single full connections with 1×1 convolution to replace the squeeze and excitation structure, and proposed a hybrid attention structure combined with spatial attention structure to enhance the ability of feature extraction. To provide multi-scale information communication capability, it constructed the multi-scale bidirectional fusion pyramid by means of bidirectional computation and multi-scale fusion. The experimental results show that the proposed method reaches 99.22% accuracy under the premise of 3.98 M parameters and 397 M floating point operations (FLOPs), and effectively identifies the combustion state in SOFC combustor. [ABSTRACT FROM AUTHOR]

Chinese spelling error correction is a task to detect and correct spelling errors in text. Most Chinese spelling errors are the misuse of semantically, phonetically or morphologically similar characters, so it is common to extract features for mode-ling different modalities. However, the direct fusion of different features or summation using fixed weights prevent the model from learning in an efficient way by ignoring the importance relationship between the information of different modalities and the bias of the model in identifying errors. This paper proposed a new model to improve this problem, called the Chinese error correction algorithm based on the fusion of text sequence error probability and Chinese spelling error probability. The method used the text sequence error probability as the dynamic weight and the common Chinese spelling error probability as the fixed weight to efficiently fuse semantic, phonetic and morphologic information. The model was able to reasonably control the inflow of different modal information into the mixed modal representation and learnt more specifically where the errors occurred. Experiments conducted on the SIGHAN benchmark show that all evaluation scores of the proposed model are improved on different datasets, which validates the feasibility of the algorithm. [ABSTRACT FROM AUTHOR]

The production and sales of high-quality fruits are inseparable from fruit quality testing, while traditional fruit quality testing methods have low precision, high cost, poor timeliness, and strong destructiveness. With the continuous advancement of science and technology in recent years, low-cost and high-efficiency non-destructive testing technology for fruit quality has been developed rapidly. Among them, hyperspectral imaging technology has gradually become a research hotspot. This paper summarizes the technical principle, application and development status of this technology in the field of fruit quality, and discusses its performance potential, existing problems, development trend and application prospects in the field of non-destructive testing of fruit quality. On the whole, hyperspectral imaging technology can non-destructively and efficiently realize the effective detection of various fruits and various quality indicators, such as maturity, sugar content, acidity, red index, etc.; Limited by hardware technology, its development focuses on data mining, that is, in the case of limited hardware development, accurate analysis results can be obtained through continuously updated and optimized targeted algorithms; on the other hand, expensive equipment, complex data processing, and poor model universality are the main problems that this technology needs to be further optimized and improved; Its future development will be based on efficient data processing of cloud computing and artificial intelligence, research and development of hyperspectral detection equipment for fruit quality with a wider range of applications, and multi-source comprehensive non-destructive testing. With the continuous development of technology, hyperspectral imaging technology has broad application prospects in non-destructive testing of fruit quality, and will become one of the important means of fruit quality testing in the future. [ABSTRACT FROM AUTHOR]

In order to solve the problems about single source of evaluation information, large deviation of accuracy and insufficient extraction fusion of heterogeneous data, this paper proposed a multi-source heterogeneous network security data fusion framework which could expand the attack behaviors. Firstly, this paper established a security event analysis model based on at-tack mode to further simplify security data. Secondly, aiming at the problem of insufficient data feature extraction at decision level, it established a 1 D-CNN model based on the attack behaviors to learn and reconstruct the features of police security data. Finally, in order to improve the classification ability in police cloud network security situation data, it modified the D-S evidence theory to fuse data combined with the credibility of multi-source security data. The experiment shows that the improved D-S theory of evidence module based on the 1 D-CNN model further improves the alarm recognition rate of security events in police cloud. Compared with other related technologies, the model has better analytical ability and a great significance for security intrusion detection and vulnerability analysis in police cloud. [ABSTRACT FROM AUTHOR]

Developers in large software development environments rely on bug reports to complete fixes. Since reporters may use different representations to describe the same bug due to different expression habits, automated detection of duplicate bug reports can reduce development redundancy as well as maintenance costs. Recent detection of repetitive bug reports tends to use deep neural networks and considers both structured and unstructured information to generate hybrid representation features. In order to obtain the features of unstructured information of bug reports more effectively, this paper proposes a D＿BBAS(Doc2 Vec and BERT BiLSTM-Attention Similarity) method, which trains a feature extraction model based on a large-scale bug report library to generate a bug summary text representation set and a bug description text representation set that can reflect deep semantic information. Then, these two distributed representation sets compute the similarity of bug report pairs, resulting in two new similarity features. These two new features will participate in the detection of duplicate bug reports when combined with the traditional features generated based on structured information. In this paper, the effectiveness of the proposed approach is verified on the bug report repositories of well-known open-source projects Eclipse, NetBeans and Open Office, which contain more than 500, 000 bug reports. The experimental results show that compared with the representative methods, the method in this paper improves the F1 value by 1.7% on average, which proves the effectiveness of the method in this paper. [ABSTRACT FROM AUTHOR]

Aiming at the problems of low detection accuracy in massive unbalanced data and poor model generalization ability caused by data heterogeneity in traditional anomaly detection methods in IoT scenarios, this paper proposed an adversarial dual-coding anomaly detection network based on federated learning(GB-AEnet-FL). Firstly, this paper proposed an active feature distribution learning algorithm based on abnormal data, which actively learns the potential feature distribution of data, expands abnormal data through data reconstruction, and balances the proportion of positive and negative samples. Secondly, this paper introduced an adversarial training mechanism in the latent feature layer adding consistency enhancement constraints and shrinkage constraints to improve the accuracy of feature extraction. Finally, this paper designed a federated learning algorithm based on dynamic model selection which compares the confidence scores of the local model and the global model, dynamically selects part of the federated bodies to participate, accelerates the aggregation of models, and protects user privacy to a certain extent. Validated on four different datasets, the results show that the proposed algorithm is better than the traditional algorithm in detection accuracy, and the generalization ability is improved accordingly. [ABSTRACT FROM AUTHOR]

Miao nationality is one of the oldest nationalities in the history of China. Miao costumes are a distinctive and colorful ethnic culture formed in the long historical process. The colorful Miao costume patterns highly summarize the pursuit of the national spirit, and the styles, colors and textures are the unique style characteristics of the costume patterns. Therefore, it is of great academic significance and application value for the digital protection and inheritance of Miao costume culture to efficiently and accurately segment minority costume patterns. At present, some scholars have carried out researches on minority costume images, but most of them are based on fuzzy C-means clustering algorithm and active contour model, and few of them are based on the deep learning algorithm for minority costume image segmentation. The traditional image segmentation algorithm used to segment the Miao costume patterns leads to the loss of some characteristics of spatial information and low efficiency and accuracy of image segmentation. In order to promote the digital protection of minority costume images and the inheritance of ethnic culture, aiming at the problems of large color differences, diversified styles and textures of Miao costume patterns, this paper proposes a Miao costume pattern segmentation model based on the attention mechanism. FCN is used as the main structure, and the feature weight of the input image is adjusted by the attention mechanism to improve the segmentation performance, so that the model can better associate the features of interest from the local level to the global level. Firstly, data enhancement is used to preprocess the image data to improve the generalization ability and robustness of the model and avoid over-fitting. Then, the feature extraction of Miao costume patterns is carried out by using the full convolutional network model fused with attention module (CBAM) to reduce the loss of spatial information, so as to effectively improve the segmentation accuracy of the model and reduce the loss rate. The model has a total of nine network layers and uses convolution, Batch Normalization (BN), CBAM, add, pooling, and concat operations. The convolution operation is used to extract image features and double the number of channels. BN operation is mainly used to normalize the training image to prevent the model from over-fitting. CBAM layers make the model pay more attention to foreground pixels while learning network weights. The add layer increases the amount of information under image features. The pooling operation performs down-sampling operation on the image to reduce the image size by two times, retain the main features while reducing the number of parameters, and improve the model generalization ability. Concat operation is used for skip connection between corresponding feature layers, so that the model can extract more rich feature information. Except Sigmoid activation function for the last convolutional layer, ReLU function is used for other layers. Analyzing the visual characteristics of ethnic costume from the perspective of deep learning and computer vision is not only convenient for researchers to store and retrieve ethnic costume images, but also conducive to the digital protection of ethnic costume images and the inheritance of ethnic culture. Finally, the experimental results on the Miao costume pattern data set show that compared with the traditional U-Net and FCN models, the IoU increases by 14.79% and 18.21%, respectively, and the Dice coefficient increases by 11.03% and 13.95%, respectively using less than half of the training parameters. Therefore, this paper provides an effective and feasible method for the research of Miao costume pattern segmentation algorithm. In the future, in-depth research regarding the mapping relationship between costume style feature points and image feature points after image segmentation will be conducted. On a certain basis, the research results play a role in the development and protection of minority costume culture, and also provide certain reference for minority costume image segmentation algorithm research. [ABSTRACT FROM AUTHOR]

Objectives: Incremental structure from motion（SfM） has become the widely used workflow for aerial triangulation（AT） of unmanned aerial vehicle （UAV） images. Recently, extensive research has been conducted to improve the efficiency, precision and scalability of SfM-based AT for UAV images. Meanwhile, deep learning-based methods have also been exploited for the geometry processing in the fields of photogrammetry and computer vision, which have been verified with large potential in the AT of UAV images. This paper aims to give a review of recent work in the SfM-based AT for UAV images. Methods:Firstly, the workflow of SfM-based AT is briefly presented in terms of feature matching and geometry solving,in which the former aims to obtain enough and accurate correspondences, and the latter attempts to solve unknown parameters. Secondly, literature review is given for feature matching and geometry solving. For feature matching, classical hand-crafted and recent learning-based methods are presented from the aspects of feature extraction, feature matching and outlier removal. For geometry solving, the principle of SfM-based AT is firstly given with some well-known and widely-used open-source SfM software. Efficiency improvement and large-scale processing are then summarized, which focus on improving the capability of SfM to process large-scale UAV images. Finally, further search is concluded from four aspects, including the change of data acquisition modes, the scalability for large-scale scenes, the development of communication and hardware, and the fusion of deep learning-based methods. Results: The review demonstrates that the existing research promotes the development of SfM-based AT towards the direction of high efficiency, high precision and high robustness, and also promotes the development of both commercial and open-source software packages. Conclusions: Considering the characteristics of UAV images, the efficiency, precision and robustness of SfMbased AT and its application need further improvement and exploitation. This paper could give an extensive conclusion and be a useful reference to the related researchers [ABSTRACT FROM AUTHOR]

In order to solve the problem of high noise and insufficient feature extraction of Retinex-Net in low-light image enhancement processing, this paper proposes a new network structure. First, the Retinex-Net network was used as the basic model to decompose the input image, and a residual shrinkage network was introduced in the convolutional layer to remove the noise generated during the decomposition process. Then, in order to preserve the details of the image and suppress noise while enhancing the brightness, the enhancement network was divided into three small sub-networks for processing respectively. Finally, the adjusted images were fused to obtain an enhanced image. Compared with the SIRE, LIME, GLADNet, Retinex-Net algorithm, experiments show that the peak signal-to-noise ratio of the images processed by the algorithm in this paper has an average increase of 3.48 dB, the mean square error has an average increase of 0.082 7, the structural similarity has an average increase of 0.146, and the lightness order error has an average increase of 271.6. [ABSTRACT FROM AUTHOR]

This paper proposes MDC image enhancement method by using joint learning of side-and central-decoding features, which considers the problems of side decoding image enhancement and central decoding image enhancement at the same time, so it can realize better network training by optimizing central decoding and side decoding features through joint learning. First, considering side independent decoding and central joint decoding features for MDC, this paper proposed a network-sharing side low-resolution feature extraction network to effectively extract features from two-side decoded images with the same content and different details, while it design a residual recursive compensation network structure and apply it into both side and central low-resolution feature extraction network. Secondly, a multiple description up-sampling reconstruction network is designed, which adopts parameter sharing strategy for partial layers of network, which can reduce parameter number of network model and improve network generalization ability. Finally, multiple description central upsampling reconstruction network is proposed to perform deep feature fusion with two low-resolution side features and central features to enhance multiple description compressed images. A large number of experimental results have shown that the proposed method is superior to several image enhancement methods such as ARCNN, FastARCNN, DnCNN, WSR and DWCNN in terms of model complexity, objective quality and visual quality assessment. [ABSTRACT FROM AUTHOR]

This paper proposed a new network framework for RGB-D salient object detection (SR-Net). In order to effectively integrate the complementarity of multi-model features, this paper took the depth feature extraction as an independent branch, use the Convolutional Block Attention Module (CBAM) to enhance the depth feature, and integrate the complementary information of the enhanced depth feature and RGB feature. Then, in order to remove feature redundancy and reduce the interference of background noise on the prediction results, it proposed a sequential refining network in the up-sampling network, that is, first, the primary global features are obtained by integrating the complementarity of multi-level and multiscale features, and used the Primary Global Feature Weight Matrix Acquisition Module (PFW) which based on the channel weight to obtains the weight matrix of the primary global feature, and then uses the obtained weight matrix to refine the features of each level to suppress the interference which caused by background noise. Finally, in order to better optimize the whole network, it proposed a new loss function. The experimental results on four public datasets show that the model is superior to nine advanced methods in different model evaluation indexes, and achieves more advanced performance. [ABSTRACT FROM AUTHOR]

Action recognition (AR) is a hot research area in computer vision field, and has an extensive application prospect for security monitoring, autopilot, production safety etc. Firstly, this paper analysed the connotation and denotation of AR and put forward the technical challenges; Secondly, the paper analysed and compared the working principles of AR from three aspects: time feature extraction, efficient optimization and long-term feature capture; Thirdly, in order to select suitable AR models for different application scenarios, this paper compared the performance characterization of 43 benchmark AR methods in recent ten years based on UCF101, HMDB51, Something-Something and Kinetics400 data sets. Finally, this paper pointed out the future development direction of AR field, and the research results can provide theoretical reference and technical support for video feature extraction and visual content understanding. [ABSTRACT FROM AUTHOR]

In order to solve the problems of large number of parameters and high GPU resource cost in binocular vision stereo matching method based on neural network in robot and unmanned driving fields, this paper proposed a multi-scale aggregation stereo matching method. Firstly, this paper proposed a multi-scale feature extraction network to obtain richer features without changing the resolution by using dilated convolution, and introduced the attention mechanism was. Then, features at different resolutions were cross-fused to improve the feature information. Secondly, the acquisition method of cost volume was changed, the cost volume was obtained by aggregation at low scale, and continuously combined the high-scale similar information to update iteratively, and cross fused multiple cost volumes to obtain the final cost volume. Finally, combined with the refinement module of attention mechanism, the outliers and discontinuous regions in the initial disparity map are corrected to obtain the final disparity map. Experimental results show that the algorithm can run under low parameter number and low cost GPU resources, and obtain good matching accuracy. [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]

In image retrieval based on deep hashing, to solve the problems of low feature extraction efficiency in convolutional neural networks (CNN ) and underutilization of feature correlation, this paper proposed a novel method combining sparse difference network and multi-supervised hashing( SDNMSH ), and used it for efficient image retrieval. SDNMSH took pairs of images as training inputs, and guided hash codes learning through an elaborately designed sparse difference convolutional neural network and a supervised hash function. The sparse difference convolutional layer and the vanilla convolutional layer composed the sparse difference convolutional neural network. The sparse difference convolutional layer could quickly extract rich feature information, to achieve efficient feature extraction of the entire network. At the same time, in order to make full use of the pairwise correlation of semantic information and features, so as to promote the feature information extracted by the network to be more effectively transformed into discriminative hash codes, and then to achieve efficient image retrieval by using SDNMSH, this paper adopted a multi-supervised hash( MSH) function and designed an objective function for this purpose. Extensive experimental results on three widely used datasets MNIST,CIFAR-10 and NUS-WIDE show that SDNMSH achieves better retrieval performance, compared with the state-of-the-arts. [ABSTRACT FROM AUTHOR]

In order to accurately predict the short-term passenger flow of urban rail transit for multiple stations, this paper proposed a deep learning model, SAE-ConvLSTM, combining convolutional long short-term memory ( ConvLSTM) and stack autoencoder (SAE) . This paper considered thirteen external factors related to passenger flow, whose features would be extracted by SAE with successive layers and thus obtain more representative features. It proposed ConvLSTM to extract spatiotemporal features of passenger flow, which was combined with the resulting external factors to predict short-term passenger flow of multiple stations simultaneously. And it developed latent action Monte Carlo tree search ( LA-MCTS) to optimize the parameters of SAE. Compared with genetic algorithm( GA), particle swarm optimization ( PSO), simulated annealing algorithm (SA) and tabu search (TS), LA-MCTS performed best in terms of effect and efficiency. This paper conducted extensive experiments. The results show that SAE-ConvLSTM works better than shallow machine learning model-back propagation neural network ( BPNN), support vector regression mode(SVR),autoregressive integrated moving average model ( ARIMA ),and deep learning modellong and short time memory network ( LSTM), convolutional neural network (CNN) and Conv LSTM without external features, ConvLSTM external features without SAE, CNN + LSTM and CNN + LSTM with external features, in terms of root mean square errors ( RMSE), mean absolute errors ( MAE) and mean absolute percentage errors ( MAPE), and the goodness of fit ( R² ) . [ABSTRACT FROM AUTHOR]