Search

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]

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]

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]

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 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]

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]

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]

Federated learning is designed to solve the problem of data fragmentation and data isolation based on privacy protection in distributed machine learning. In the federated learning system, participants collaboratively train a model. Each participant uses local data to train the local model, and uploads the trained local model to the server for aggregation. In the real application environment, the data distribution between nodes is often very different, resulting in the accuracy of federated learning model is low. In order to solve the influence of non-independent identically distributed data on the accuracy of the model, this paper proposed a clustering federated learning framework by using the similarity of data distribution between different nodes. Extensive experiments were conducted on Synthetic, CIFAR-10 and FEMNIST standard datasets. Compared with other federated learning methods, clustering federated learning based on data distribution greatly improves the accuracy of the model and requires less computation. [ABSTRACT FROM AUTHOR]

There is a huge demand for weather cloud images in the observatory' s weather monitoring system. In order to solve the problems of model instability and loss of image features when the conventional generative adversarial network expands the dataset of the weather cloud images, this paper proposed a double-layer embedded adversarial image generation method based on SAU-Net DCGAN. This method consisted of two layers of networks which were nested within each other. Firstly,by the first layer of embedded network, it added the U-shaped network to the generator of the generative adversarial network. This network acted as the basic architecture and enhanced the feature recovery capability of the image by using the jump connection between the encoder and the decoder. Secondly,by the second layer of embedded network, it added SA to the U-shaped network. This attention mechanism reduced the complexity of the model by simplifying the parameters, improved effectively the feature loss in the dark part of the image. Finally, it developed a new weight calculation method to strengthen the connection between each features and improved the extraction of detail texture features from the images. The experimental results show that the quality of the images generated by this method is better than that of the conventional generative adversarial network in terms of sharpness and saturation. The evaluation indicators PSNR and SSIM have increased by 27. 06 dB and 0. 606 5 respectively. [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]

Most of the existing steganalysis methods have weak generalization ability and cannot effectively detect unknown steganalysis algorithms, which makes the accuracy of their classification greatly reduced in the practical application process. To solve this problem, this paper proposed an image steganalysis method based on group convolution and snapshot ensembling（Snapshot Ensembling steganalysis network, SENet）. Firstly, residual convolution block and group convolution block are used to extract the features of the image. Secondly, the model with the best performance is obtained as the snapshot model in each training period. Finally, the selected snapshot models are integrated to classify the images. This method uses the techniques of group convolution and snapshot ensembling to avoid the high training cost of traditional integration methods and the limited generalization ability of a single classifier. Experimental results show that this method can improve the accuracy of steganographic analysis model, and can effectively classify when the training set and test set are mismatched, and has high model generalization ability. [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]

Due to ignoring the extraction of multi-scale features and the differences between features at different levels, salient object detection still has the problems of incomplete prediction structure and loss of details. To solve these problems, this paper proposes a new saliency detection model M3 Net. The network mainly consists of Multi-scale Features Adaptive Fusion Module (MAF) and Recurrent Feedback Aggregation Module (RFA). MAF can adaptively capture and aggregate multi-scale features at different levels. RFA can effectively prevent feature dilution while aggregating features at different levels in the iterative process. The experimental results on 5 benchmark datasets show that the network outperforms 10 existing networks in three evaluation metrics: Fβ, Em, and MAE. On the DUT-OMRON dataset, the Fβ is 0.4% higher than the second-ranked saliency detection model, and the Em is 0.3% higher. In ECSSD dataset, the Fβ is 0.2% higher than the second-ranked saliency detection model, and the Em is 0.3% higher, and the network also has excellent speed performance [ABSTRACT FROM AUTHOR]

The sentence level extraction method is insufficient to solve the problem of Chinese event element dispersion. To solve this problem, this paper proposed a document level event extraction method based on context fusion. Firstly, the paper divided the document into paragraphs, and used bidirectional long and short memory network to extract sequence features of paragraphs. Secondly, the method used self-attention mechanism to capture the interaction information of paragraph context. Then the method combined the document sequence features with the interaction information to update the semantic representation. Finally, the method used sequence annotation to extract event elements and match event types. Compared with other event extraction methods on the same Chinese data set, the experimental results show that this method can effectively extract scattered event elements from documents, and improve the extraction performance of the model. [ABSTRACT FROM AUTHOR]

Point clouds are a large collection of points and have important geometric structures. Due to its large amount of data, it is unavoidable that some similar points will appear in some areas, which makes the feature extraction to extract some duplicate information, resulting in redundancy of calculation, and reducing the accuracy of training. To solve these problems, this paper proposed a new neural network—PointPCA, which could effectively solve the above problems. In PointPCA, there were three modules: a) sampling module. It proposed an average point sampling (APS) sampling method, which could effectively avoid some similar points and get a new set of points that approximate this set of point clouds. b) feature extraction module. It used the idea of grouping to extraction multi-scale spatial feature for the set of new points. c) stitching module. It stitched the feature vectors extracted at each scale into one feature vector. Experiments show that the accuracy of PointPCA is 4.6% higher than that of PointNet and 1.1% higher than that of PointNet++. It also works well in mIoU assessment tests. [ABSTRACT FROM AUTHOR]

In order to solve the problem that the Transformer model cannot apply to mobile devices with limited computational resources due to its huge number of parameters and computational complexity, this paper proposed a mobile-friendly lightweight image recognition network called FuseNet. FuseNet utilized convolutional neural network to extract local feature information and self-attention to excel in modeling global information, and it integrated the features of both local and global representations into a single feature extraction module, which efficiently combined the advantages of the two different structures to achieve a high accuracy with a small model size. Experiments demonstrate that FuseNet with different model sizes can achieve good performance without using pre-training and it can well apply to mobile devices. FuseNet-B achieves 80. 5% accuracy with 14. 8 M parameters on ImageNet-1 K dataset, the performance exceeds the same volume of Transformer models and convolutional neural networks. [ABSTRACT FROM AUTHOR]

In the field of CXR image classification, current studies mostly use convolutional neural network and transfer learning technology. But in the process of rapidly constructing a network without considering the specificity of the CXR image. To solve the above problem, this paper proposed a novel XDense-RC-net. Based on the DenseNet model, this method improved the capability of feature extraction and feature fusion by introducing a new spatial attention mechanism in the densely connected layer, Optimized Transition module of DenseNet by using two different pooling strategies to enhance the noise immunity of the model. XDense-RC-net used the chest X-ray14 (multilabel and 14-class) and COVIDx (single label and 3-class) datasets for validating. In the multi-label classification experiments, the average AUC score reached 0.854, which is 0.109 higher than the benchmark method. In the single-label classification experiments, the average accuracy reached 96.75%, which is 7.75% higher than the baseline model. The results show that XDense-RC-net improves the accuracy of CXR image classification and can generalize to multilabel and single-label classification tasks. [ABSTRACT FROM AUTHOR]

The current ADMET classification methods have shortcomings in classifying the ADMET of compounds data with multiple characteristics and feature correlation. Moreover, the classification results of ADMET are not explainable. To solve these problems, this paper proposed a classification model based on capsule network（CapsMC). CapsMC model first proposed the Feature-to-Image algorithm. It used this algorithm to consider the correlation and dependence between features into the classification basis, and realized the multi-level extraction of features. Second, it explored the advanced application of capsule network, and designed a cognitive reasoning mechanism. It used this mechanism to carry out cognitive reasoning on features, and realized explainable classification of ADMET. Experimental results on five ADMET datasets show that CapsMC can achieve the explainable classification of ADMET well. [ABSTRACT FROM AUTHOR]

Most of the eye disease classification methods are to classify a single category of diseases at different levels, and the network model has problems such as large parameters and complex calculations. To solve these problems, this paper proposed a lightweight fundus disease multi-classification network called MELCNet. The network used PPLCNet as the backbone, which was composed of input layer for feature extraction, parallel multi-scale structure, bilinear SE attention module, depth separable convolution and the h-swish activation function which calculated with smaller parameters, in order to pay attention to the key disease information of different scales and different diseases. Experimental results show that the proposed multi-scale attention lightweight network model has less parameters and computational complexity, and excellent classification results have been obtained in the multi-classification of the four selected fundus diseases and normal fundus images. The classification accuracy in the internal combined test set is 1.11 % higher than that in ResNet-50,and the performance on the public datasets is improved by 2. 5% relative to the multi-classification literature of similar type network by Xie et al, and compared with other lightweight classification networks, it has higher accuracy and strong robustness in the field of multi-classification of fund us diseases. [ABSTRACT FROM AUTHOR]

Aiming at the problem of missing detection and false detection caused by inaccurate motion feature extraction of existing video key frame extraction algorithms, this paper proposed a reinforcement learning key frame extraction algorithm combining multi-channel feature and attention mechanism. The algorithm extracted the human skeleton joint points from the video sequence through the human posture recognition algorithm firstly. Then it used the S-GCN and ResNet50 network to extract the motion features and static features in the video sequence respectively, and performed a weighted fusion of the two. Finally, it applied the attention mechanism to calculate the importance of the video frame of the feature sequence, and used reinforcement learning to extract and optimize key frames. The experimental results show that the algorithm can solve the problem of missing and false detection in the key frame extraction of motion video. It performs well in the detection of video frames containing key actions, with high accuracy and strong stability. [ABSTRACT FROM AUTHOR]

Existing deep learning based single image super-resolution (SISR) models usually improve the fitting ability of the model by increasing the number of network layers, but fail to fully extract and reuse features, leading low quality of reconstructed images. To solve this problem, this paper proposed an image super-resolution model based on feature fusion and attention mechanism. This model used residual in residual (RIR) structure in feature extraction module. The feature extraction module of the network consisted of several residual groups. Each residual group consisted of several residual blocks. This module implemented local feature fusion in each residual group and global feature fusion between each group. In addition, this model introduced coordinate attention module into each residual block and spatial attention module into each residual group. It verifies that the model is able to fully extract features and reuse features. The final experimental results show that the model is superior to the existing models in objective evaluation indexes and subjective visual effect. [ABSTRACT FROM AUTHOR]

Deep convolutional neural network is widely used in the field of medical image segmentation. In recent years, multi-scale fusion structure is used to improve segmentation network, which often increases the complexity of the model and reduces the training efficiency while improving the accuracy. To solve these problems, this paper proposed a novel segmentation algorithm for WU-Net pulmonary nodules. It improved the U-Net by introducing an improved residual connection module into the original down-sampling coding channel and mean while using the information channel improved by the new dep module to complete feature extraction and feature fusion. In the experiment, it used LUNA16 dataset to train and verify the model. Meanwhile, comparative experiments used some newly segmentation models. In the experiment with scale of nodules, the Dice coefficient and IoU of proposed model can reach 96.72% and 91.78% respectively.F1-score can reach 92.41% after adding 10% negative samples, which is 1.23% higher than UNet3+.In the experiment with scale of lung parenchyma, the Dice coefficient and IoU of proposed model can reach 83.33% and 66.79% respectively, which is 1.35% and 2.53% higher than RU-Net. The training efficiency of WU-Net is the highest, which is 39.6% higher than U-Net spend. The result shows that WU-Net has improved segmentation effects and training efficiency of model as well. [ABSTRACT FROM AUTHOR]

To solve the problem of category-oriented 3D articulated objects pose estimation, this paper proposed a category-oriented 3D object pose estimation method based on key points. It designed a network based on key points, which used PointNet + + as base network and estimated the pose of the object through feature extraction, key points extraction, and pose prediction with high accuracy and robust. At the same time, based on ANCSH, it proposed a hierarchical standardized category representation method of key points for K-AOPE, which only required a few points to describe the object category. To evaluate the method, it did some experiments on the public motion dataset shape2motion. The results show that the rotation angle errors of the proposed method are 2. 3°, 3. 1°, 3. 7°, the translation errors are 0. 034, 0. 030, 0. 046, the joint state error are 2. 4°, 2. 5°, and the joint parameter error are 1. 2°, 0. 9°, 0. 008, 0. 010. Compared with other similar methods, the proposed method has lower computational complexity and higher robust. [ABSTRACT FROM AUTHOR]

To solve the problems of difficult initialization and low positioning accuracy of SLAM algorithm caused by sparse feature points and fast camera movement in practical application scenes, this paper proposed a visual-inertial SLAM algorithm. It used EDLines feature extraction and matching algorithm to introduce rich line features to improve the robustness of feature tracking. Using the complementary advantages of the inertial sensor and the visual sensor, it carried out the optimization in three steps in the initial stage through the visual-inertial initialization by steps method, so as to improve the accuracy and speed of the initialization. Experiments show that the line feature extraction and matching algorithm used in this algorithm provides faster matching speed and lower mismatching rate than the traditional LSD algorithm. Moreover, the proposed algorithm has higher accuracy and stronger robustness in the estimation of camera pose in the visual-inertial data set. [ABSTRACT FROM AUTHOR]

The widespread use of social networking platforms has led to the emergence of emotionally rich online comment texts, analyzing the emotions expressed in comments is of great significance to companies, platforms, etc. In order to solve the current problem of weak feature extraction ability and ignoring the emotional information of short text in online comment short text sentiment analysis, this paper proposed a model based on text sentiment value weighted char-word mixture word representation-SVW-BERT. First, it based on the fusion of character and word level vectors represented text vectors for maximizing semantic representation. At the same time, considering the influence of adverbs, negative words, exclamation sentences and interrogative sentences on the sentiment of the text, it used the weight to calculate the sentiment value of the text, and constructed sentiment analysis model of Chinese short text based on text sentiment value weighted char-word mixture word representation. Through the network platform online reviews data set, it validated the feasibility and the advantages of the model. The experimental results show that the char-word mixture word representation is stronger in semantic extraction, and the sentiment value weighted sentence vector considers the sentiment information contained in the text itself, which achieves the effect of improving the ability of sentiment classification. [ABSTRACT FROM AUTHOR]

The single-frame image super-resolution（SISR） method based on deep networks is currently a hotspot in SR research. However, most of these methods mainly focus on exploring the network depth structure during feature extraction, ignoring the similarity between the feature layers in the intermediate space, and neglect the feature differences between the feature layers during reconstruction. This paper proposed a progressive network based on spatial feature transformation and back-projection reconstruction to solve this problem. The main feature of this method was to perform feature affine transformation on the feature space when extracting image features to increase the different similarities between feature layers, so as to obtain progressive features and spatial transformation features. In the image reconstruction stage, the reconstruction module adopted the strategy of multi-scale back projection to integrate the multi-source features of the image, which made the module pay more attention to the difference between features. Finally, the experimental results show that compared with most super-resolution methods, the proposed method has greatly improved the evaluation indexes such as PSNR/SSIM during image super-resolution reconstruction, and the texture information of the reconstructed image is more abundant. [ABSTRACT FROM AUTHOR]

The size of the receptive field of the convolutional neural network is related to the size of the convolution kernel. And the traditional convolution uses a fixed-size convolution kernel, which limits the feature perception ability of the network model. In addition, due to the parameter sharing mechanism of the convolutional neural network, it used the same feature extraction method for all pixels in the spatial region. However, there are differences in the distribution of noise signals and clean speech signals in the noisy spectrogram, especially in the complex noise environment, the general convolution method is difficult to achieve high-quality speech signal feature extraction and choosing. In order to solve the above problems, this paper proposed a multi-scale region adaptive convolution module, which used multi-scale information to improve the feature perception ability of the model and automatically allocated the area adaptive convolution achieve and improved the denoising ability of the model. The experiments on the TIMIT public datasets show that the proposed algorithm has achieved satisfactory results in the metrics of speech quality and intelligibility. [ABSTRACT FROM AUTHOR]

In order to solve the problems of insufficient feature extraction of sEMG end-to-end gesture recognition and low accuracy of multi-gesture recognition, this paper proposed a multi-stream convolutional EMG gesture recognition network model fusing attention mechanism. The proposed model generated sub-myoelectrograms from multi-channel time-domain sEMG through sliding windows and used a multi-stream convolutional neural network to fully extract the semantic features of each sampled channel sEMG, and then aggregated them to obtain rich multi-channel gesture semantic features. At the same time, the proposed model calculated the attention distribution map of semantic features from temporal and feature channel dimensions to strengthen the useful features and weaken the useless features, further improved the accuracy of multi-gesture recognition. The experiment used the Ninapro dataset for training and testing and compared with the mainstream electromyogram recognition model. The experimental results show that the proposed model has better performance in recognition accuracy, which proves the effectiveness of the proposed model. [ABSTRACT FROM AUTHOR]

Capsule network emphasizes the encoding of image features' spatial relationship, while its feature extraction module has difficulty in handling complex scenarios. In order to improve the performance of CapsNet, this paper proposed a capsule network with a self-attention feature extraction module ( SA-CapsNet) to solve the above problem. At first, it improved CapsNet by reducing the dimension of the capsule and adding an intermediate layer. Then mapped the SA module to the feature extraction layer of the capsule network to increase feature extraction capability. The results worked on the MNIST, Fashion MNIST, and CIFARlO datasets show that the classification accuracy rates are 99. 67%, 92. 21 %, and 82. 51 %, respectively. Additionally, results demonstrate the effectiveness of this proposed network, and the overall performance has been dramatically improved compared with CapsNet [ABSTRACT FROM AUTHOR]