Feature Representation Ability Research Articles

A multiple gated boosting network for multi‐organ medical image segmentation

AbstractSegmentations provide important clues for diagnosing diseases. U‐shaped neural networks with skip connections have become one of popular frameworks for medical image segmentation. Skip connections really reduce loss of spatial details caused by down‐sampling, but they cannot handle well semantic gaps between low‐ and high‐level features. It is quite challenging to accurately separate out long, narrow, and small organs from human bodies. To solve these problems, the authors propose a Multiple Gated Boosting Network (MGB‐Net). To boost spatial accuracy, the authors first adopt Gated Recurrent Units (GRU) to design multiple Gated Skip Connections (GSC) at different levels, which efficiently reduce the semantic gap between the shallow and deep features. The Update and Reset gates of GRUs enhance features beneficial to segmentation and suppress information adverse to final results in a recurrent way. To obtain more scale invariances, the authors propose a module of Multi‐scale Weighted Channel Attention (MWCA). The module first uses convolutions with different kernel sizes and group numbers to generate multi‐scale features, and then adopts learnable weights to emphasize the importance of each scale for capturing attention features. Blocks of Transformer Self‐Attention (TSA) are sequentially stacked to extract long‐range dependency features. To effectively fuse and boost the features of MWCA and TSA, the authors use GRUs again to propose a Gated Dual Attention module (GDA), which enhances beneficial features and suppresses adverse information in a gated learning way. Experiments show that the authors’ method achieves an average Dice coefficient of 80.66% on the Synapse multi‐organ segmentation dataset. The authors’ method outperforms the state‐of‐the‐art methods on medical images. In addition, the authors’ method achieves a Dice segmentation accuracy of 62.77% on difficult objects such as pancreas, significantly exceeding the current average accuracy, so multiple gated boosting (MGB) methods are reliably effective for improving the ability of feature representations. The authors’ code is publicly available at https://github.com/DAgalaxy/MGB‐Net.

γ-polyglutamic acid fermentation monitoring with ATR-FTIR spectroscopy based on a shallow convolutional neural network combined with data augmentation and attention module

The potential of attenuated total reflection Fourier transform infrared (ATR-FTIR) spectroscopy to online real-time monitor the γ-polyglutamic acid (γ-PGA) fermentation process by measuring the concentration of two key components, glucose and sodium glutamate, was investigated. Partial least squares regression (PLSR) and convolutional neural network (CNN) were selected as the multivariate calibration model to predict fermentation parameters. To solve the small sample training problem with CNN, the data augmentation strategy was firstly combined with a shallow CNN to generate a new CNN model termed as DA-CNN. To improve the performance of DA-CNN further, squeeze-and-excitation (SE) module of the attention mechanism was integrated into DA-CNN to form a DA-CNN variant termed as DA-SE-CNN, which could capture the different importance of extracted channel-wise feature to improve feature representation ability. Experiments were conducted on the γ-PGA dataset, and the results showed that the nonlinear models (CNN, DA-CNN and DA-SE-CNN) performed better than the linear method (PLSR) in all cases except the case of sodium glutamate prediction with CNN. Additionally, both DA-CNN and DA-SE-CNN perform better than CNN in all cases with limited training data. Especially DA-SE-CNN gave the best result with excellent prediction accuracy, which indicated that ATR-FTIR combined with nonlinear regression tool (DA-SE-CNN) as a rapid method to monitor γ-PGA fermentation process is feasible.

FTransCNN: Fusing Transformer and a CNN based on fuzzy logic for uncertain medical image segmentation

The accurate segmentation of medical images plays a crucial role in diagnosing and treating diseases. Although many methods now use multimodal joint segmentation, the joint use of segmentation features extracted by multiple models can lead to heterogeneity and uncertainty. Unreasonable fusion methods cannot exploit the advantages of multiple models and still lack good performance in segmentation. Therefore, this study proposes the FTransCNN model, which is composed of a convolutional neural network (CNN) and Transformer and is based on a fuzzy fusion strategy that jointly utilizes the features extracted by a CNN and Transformer through a new fuzzy fusion module. First, the CNN and Transformer act as the backbone network for parallel feature extraction. Second, channel attention is used to promote the global key information of Transformer to improve the feature representation ability, and spatial attention is used to enhance the local details of CNN features and suppress irrelevant regions. Third, the proposed model applies the Hadamard product to model fine-grained interactions between the two branches and uses the Choquet fuzzy integral to suppress heterogeneity and uncertainty in fused features. Fourth, FTransCNN employs fuzzy attention fusion module (FAFM) hierarchical upsampling to effectively capture both low-level spatial features and high-level semantic context. Finally, the new model obtains the final segmentation result by using the deconvolution and results in an improvement in segmentation. The experimental results on Chest X-ray and Kvasir-SEG dataset show that FTransCNN has better performance on segmentation tasks than the-state-of-the-art deep segmentation models.

DRnet: Dynamic Retraining for Malicious Traffic Small-Sample Incremental Learning

Deep learning has achieved good classification results in the field of traffic classification in recent years due to its good feature representation ability. However, the existing traffic classification technology cannot meet the requirements for the incremental learning of tasks in online scenarios. In addition, due to the high concealment and fast update speed of malicious traffic, the number of labeled samples that can be captured is scarce, and small samples cannot drive neural network training, resulting in poor performance of the classification model. Therefore, this paper proposes an incremental learning method for small-sample malicious traffic classification. The method uses the pruning strategy to find the redundant network structure and dynamically allocates redundant neurons for training based on the proposed measurement method according to the difficulty of the new class. This enables the network to perform incremental learning without excessively consuming storage and computing resources, and reasonable allocation improves the classification accuracy of new classes. At the same time, through the knowledge transfer method, the model can reduce the catastrophic forgetting of the old class, relieve the pressure of training large parameters with small-sample data, and improve the model classification performance. Experiments involving multiple datasets and settings show that our method is superior to the established baseline in terms of classification accuracy, consuming 50% less memory.

DGCN-ED: dynamic graph convolutional networks with encoder–decoder structure and its application for airborne LiDAR point classification

ABSTRACT Achieving high-precision automatic classification in real-world applications for airborne laser scanning point clouds is a challenging task duo to their unstructured nature, uneven density distribution, high redundancy, incompleteness and scene complexity. Graph Convolutional Neural Networks can process scattered point clouds directly without regularization, which avoids the loss of depth information，and has recently become a topic of increased interest. Therefore, this study proposed an extension of the Graph-Unet network architecture named DGCN-ED for airborne LiDAR point classification, which uses a Graph Convolutional Neural Network as a representation to describe complex object relationships and an encoder-decoder architecture to capture the multi-scaled point features and describe objects in the high-level feature space. The two-layer dynamic update Graph Convolutional Neural Network is designed to expand the effective range of nodes and enhance the representation ability of learned pointwise features. The effectiveness of the proposed method is evaluated by an experiment on the ISPRS Vaihingen 3D semantic labelling benchmark dataset. Moreover, experiments on the IEEE 2019 Data Fusion Contest Dataset were conducted to demonstrate the generalization abilities of the proposed method. The results show that our method achieved on average 2.8% higher overall accuracy than existing methods, with an overall accuracy of 98.0% and an average F1 score of 0.797.

Pulmonary nodule detection based on Hierarchical-Split HRNet and feature pyramid network with atrous convolution

Accurate pulmonary nodule detection is crucial to the diagnosis of lung diseases. In this work, we propose an end-to-end network for pulmonary nodule detection mainly consisting of pre-processing, detection modules for candidate prediction, and a discriminator to identify the existence of nodules. In the detection module, HS-HRNet is proposed to fulfill feature extraction on high-resolution input for pulmonary nodules that occupy tiny spaces of CT images. HS-HRNet incorporates plug-and-play Hierarchical-Split block into High-Resolution Network (HRNet) and modifies STEM with sandglass module. The main advantages of these modifications are that HS-HRNet can largely promote feature representation ability by split and concatenation operation with no significant increase in computation. In addition, a novel Feature Pyramid Network with Atrous Convolution (AC-FPN) is proposed for multi-scale feature fusion and multi-level prediction. This design allows context and spatial information in different feature levels to be leveraged and extracted under larger receptive fields. Besides, a discriminator replaces false positive reduction modules in most pulmonary detection methods. The discriminator judges the existence of nodules and back-propagates prediction proposals to detection modules through adversarial training. Experiments on publicly available datasets demonstrate competitive performance in pulmonary nodule detection.

Enriched multi-scale cascade pyramid features and guided context attention network for industrial surface defect detection

Surface defect detection is a very important technique to guarantee product quality in industrial fields. However, the detection of multi-scale defects and defects with poor visibility is still a challenging problem. To address this issue, we propose a novel network by collaborating multi-scale cascade pyramid features and a guided context attention mechanism for the pixel-wise defection of surface defects, called MPA-Net. The MPA-Net is a full y-convolutional network (FCN) with an encoder–decoder architecture, which can integrate multi-scale features and merge them into the different stages of the decoder for generating the defect segmentation map. Specifically, the proposed guided context attention module (GCA) is used to transmit the global context information from the large scale to the small scale, which can promote the initial recovery capability of the decoder, and thus help to locate defects with different sizes and defects with poor visibility. Moreover, the proposed pyramid feature fusion and enrichment module (FFEM) is employed to aggregate low-level coarse features and high-level semantic features in each scale, so as to increase the ability of defect feature representation. The aggregation features at different scales are then fused to the different layers of the decoder, which is beneficial to recover the details of defects gradually. The evaluation results on four public datasets demonstrate that the proposed method has excellent performances on mean intersection of union (DAGM2007: 64.94%, KolektorSSD: 77.90%, RSDDs-I: 86.63%, RSDDs-II: 80.62%, FID: 96.98%) and mean pixel accuracy (DAGM2007: 67.97%, KolektorSSD: 85.01%, RSDDs-I: 94.13%, RSDDs-II: 88.53%, FID: 98.71%).

A cross-view geo-localization method guided by relation-aware global attention

Cross-view geo-localization mainly exploits query images to match images from the same geographical location from different platforms. Most existing methods fail to adequately consider the effect of image structural information on cross-view geo-localization, resulting in the extracted features can not fully characterize the image, which affects the localization accuracy. Based on this, this paper proposes a cross-view geo-localization method guided by relation-aware global attention, which can capture the rich global structural information by perfectly integrating attention mechanism and feature extraction network, thus improving the representation ability of features. Meanwhile, considering the important role of semantic and context information in geo-localization, a joint training structure with parallel global branch and local branch is designed to fully mine multi-scale context features for image matching, which can further improve the accuracy of cross-view geo-localization. The quantitative and qualitative experimental results on University-1652, CVUSA, and CVACT datasets show that the algorithm in this paper outperforms other advanced methods in recall accuracy (Recall) and image retrieval average precision (AP).

Intelligent fault diagnosis of partial deep transfer based on multi-representation structural intraclass compact and double-aligned domain adaptation

The main challenge of fault diagnosis models based on partial domain adaptation (PDA) is to promote positive transfer in the shared label space and avoid negative transfer caused by the mismatch between the outlier and the target label spaces. To address the above challenges and enhance the diagnosis performance in the PDA scenario, this paper proposes a partial deep transfer diagnosis model (MICDDA) based on multi-representation structure intraclass compact and double-aligned domain adaptation. First, the Gabor feature extractor and multi-representation structure network are constructed to enhance the representation ability of deep features and enrich feature information. Then, the information entropy and class diversity loss of the target domain are introduced to improve the loss of the classification network, and the source class-level weight and the target instance-level weight are used to reduce the outlier features participating in the alignment of the target features. Next, an intraclass compact constraint is proposed to enhance the agglomeration of features with the same label by minimizing the distance between the feature and the centroid. Finally, a double-aligned domain adaptation strategy is designed to enhance the alignment between the source and target domains' shared features by minimizing the inter- and intra-domain alignment. Experimental results in two case studies show that the MICDDA can effectively weaken the transfer of source outlier knowledge to the target domain, capture rich and homogeneous compact fault features in PDA diagnosis scenarios, and effectively improve diagnostic performance.

Driver Identification Using Deep Generative Model With Limited Data

The scarcity of driving data constrains the accuracy of deep learning (DL)-based driver identification methods in practical application scenarios. To address this issue, this study proposes a novel unsupervised deep generative model called the convolution condition variant autoencoder (CCVAE) for driving data augmentation. In CCVAE, aided by driver identification information, the condition variant autoencoder can learn the real driving data distribution of each driver through an unsupervised learning paradigm; and aiming for better feature representation ability, convolutional neural network and deconvolution are leveraged, respectively. Therefore, a large number of synthetic samples can be generated by the generative part of the CCVAE. We demonstrate the effectiveness of the CCVAE through extensive experimental analysis using a real dataset collected from a vehicular CAN bus; the improvement of the DL-based driver identification results is demonstrated using synthetic samples. For instance, when only using 2% of the original data, approximately 20% improvement is achieved in terms of four evaluation indicators for two commonly used DL-based driver identification methods, namely, 1-D CNN and LSTM. Furthermore, several comparable experiments with state-of-the-art deep generative methods reveal the superior performance of the proposed CCVAE with respect to identification results, synthetic data quality, and model computation time. Therefore, the proposed model accomplishes a breakthrough in driver identification with limited data and shows great potential in data-driven applications of intelligent vehicles.

Semi-supervised medical image segmentation using adversarial consistency learning and dynamic convolution network.

Popular semi-supervised medical image segmentation networks often suffer from error supervision from unlabeled data since they usually use consistency learning under different data perturbations to regularize model training. These networks ignore the relationship between labeled and unlabeled data, and only compute single pixel-level consistency leading to uncertain prediction results. Besides, these networks often require a large number of parameters since their backbone networks are designed depending on supervised image segmentation tasks. Moreover, these networks often face a high over-fitting risk since a small number of training samples are popular for semi-supervised image segmentation. To address the above problems, in this paper, we propose a novel adversarial self-ensembling network using dynamic convolution (ASE-Net) for semi-supervised medical image segmentation. First, we use an adversarial consistency training strategy (ACTS) that employs two discriminators based on consistency learning to obtain prior relationships between labeled and unlabeled data. The ACTS can simultaneously compute pixel-level and image-level consistency of unlabeled data under different data perturbations to improve the prediction quality of labels. Second, we design a dynamic convolution-based bidirectional attention component (DyBAC) that can be embedded in any segmentation network, aiming at adaptively adjusting the weights of ASE-Net based on the structural information of input samples. This component effectively improves the feature representation ability of ASE-Net and reduces the overfitting risk of the network. The proposed ASE-Net has been extensively tested on three publicly available datasets, and experiments indicate that ASE-Net is superior to state-of-the-art networks, and reduces computational costs and memory overhead. The code is available at: https://github.com/SUST-reynole/ASE-Net.

M2FINet: Modality-specific and Modality-shared Features Interaction Network for RGB-IR Person Re-Identification

Learning rich and discriminative person-related modality-shared feature representations to distinguish the same person in different modalities is significant for RGB-Infrared Person Re-IDentification (RGB-IR ReID). However, most existing models often directly extract modality-shared features from those modality-specific features without considering their interactions. This will result in the ineffective exploration of person-related modality-shared features. To address such a problem, a novel Modality-specific and Modality-shared Features Interaction Network (M2FINet) model is proposed for RGB-IR ReID in this paper. Especially, in the proposed M2FINet model, a Cross-level Feature Guidance and Injection (CFGI) module is carefully designed to establish and exploit the interactions between the middle-level modality-specific features and the high-level modality-shared features. Specifically, the proposed CFGI module mainly consists of two streams, a shared-to-specific feature guidance (H2P) stream and a specific-to-shared feature injection (P2H) stream. Among that, the H2P stream aims to take high-level modality-shared features as the prior information to guide the re-exploration of more rich and discriminative modality-shared semantic information from such middle-level modality-specific features. The P2H stream aims to enhance the representation ability and discriminability of high-level modality-shared features by introducing more modality-shared detail information from the middle-level modality-specific features. On top of that, a simple but effective feature aggregation module, i.e., Focusing on Person (FOP), is further designed in our proposed model to reinforce such discriminative modality-shared features within those person-related regions via a multi-pooling feature aggregation manner. Extensive experiments on two public benchmarks, i.e, SYSU-MM01 and RegDB, show that our proposed model consistently improves the accuracy of RGB-IR ReID. Without bells and whistles, it achieves Rank-1/mAP by 74.73%/68.96% on the large-scale SYSU-MM01 dataset.

3D ground penetrating radar road underground target identification algorithm using time-frequency statistical features of data

3D ground penetrating radar is the main detection technology used to locate underground targets on urban roads, and the current underground target intelligent identification algorithms for this technology rely on two-dimensional images. Relying only on radar image texture as the identification feature, due to the limitations of single dimensionality and single feature utilization, leads to significant errors of these algorithms in identifying underground targets. For the above problems, an underground target depth time-frequency statistical feature identification algorithm based on 3D convolution using raw radar data is proposed. The detailed changes and overall trends of underground targets are extracted by multi-size 3D receptive fields. Meanwhile, the depth time-frequency statistical features of multiple position slices in the horizontal, vertical, crossed-vertical, left-right diagonal, and front-back diagonal directions of radar data are constructed at different layers. Multi-dimensional time-frequency statistical features are input to a fully connected classifier, and the final result is obtained by combining the classification results of the seven directions through the Bagging mechanism. This algorithm expands the feature dimension and direction of judging underground targets and strengthens the feature representation ability. It makes full use of the 3D ground penetrating radar data information and reflects the features of underground space from multiple angles, which improves the identification accuracy of underground targets.

Detection of Chrysanthemums Inflorescence Based on Improved CR-YOLOv5s Algorithm.

Accurate recognition of the flowering stage is a prerequisite for flower yield estimation. In order to improve the recognition accuracy based on the complex image background, such as flowers partially covered by leaves and flowers with insignificant differences in various fluorescence, this paper proposed an improved CR-YOLOv5s to recognize flower buds and blooms for chrysanthemums by emphasizing feature representation through an attention mechanism. The coordinate attention mechanism module has been introduced to the backbone of the YOLOv5s so that the network can pay more attention to chrysanthemum flowers, thereby improving detection accuracy and robustness. Specifically, we replaced the convolution blocks in the backbone network of YOLOv5s with the convolution blocks from the RepVGG block structure to improve the feature representation ability of YOLOv5s through a multi-branch structure, further improving the accuracy and robustness of detection. The results showed that the average accuracy of the improved CR-YOLOv5s was as high as 93.9%, which is 4.5% better than that of normal YOLOv5s. This research provides the basis for the automatic picking and grading of flowers, as well as a decision-making basis for estimating flower yield.

DrugormerDTI: Drug Graphormer for drug–target interaction prediction

Drug-target interactions (DTI) prediction is a crucial task in drug discovery. Existing computational methods accelerate the drug discovery in this respect. However, most of them suffer from low feature representation ability, significantly affecting the predictive performance. To address the problem, we propose a novel neural network architecture named DrugormerDTI, which uses Graph Transformer to learn both sequential and topological information through the input molecule graph and Resudual2vec to learn the underlying relation between residues from proteins. By conducting ablation experiments, we verify the importance of each part of the DrugormerDTI. We also demonstrate the good feature extraction and expression capabilities of our model via comparing the mapping results of the attention layer and molecular docking results. Experimental results show that our proposed model performs better than baseline methods on four benchmarks. We demonstrate that the introduction of Graph Transformer and the design of residue are appropriate for drug–target prediction.

CTformer: convolution-free Token2Token dilated vision transformer for low-dose CT denoising

Objective. Low-dose computed tomography (LDCT) denoising is an important problem in CT research. Compared to the normal dose CT, LDCT images are subjected to severe noise and artifacts. Recently in many studies, vision transformers have shown superior feature representation ability over the convolutional neural networks (CNNs). However, unlike CNNs, the potential of vision transformers in LDCT denoising was little explored so far. Our paper aims to further explore the power of transformer for the LDCT denoising problem. Approach. In this paper, we propose a Convolution-free Token2Token Dilated Vision Transformer (CTformer) for LDCT denoising. The CTformer uses a more powerful token rearrangement to encompass local contextual information and thus avoids convolution. It also dilates and shifts feature maps to capture longer-range interaction. We interpret the CTformer by statically inspecting patterns of its internal attention maps and dynamically tracing the hierarchical attention flow with an explanatory graph. Furthermore, overlapped inference mechanism is employed to effectively eliminate the boundary artifacts that are common for encoder-decoder-based denoising models. Main results. Experimental results on Mayo dataset suggest that the CTformer outperforms the state-of-the-art denoising methods with a low computational overhead. Significance. The proposed model delivers excellent denoising performance on LDCT. Moreover, low computational cost and interpretability make the CTformer promising for clinical applications.

MCANet: shared-weight-based MultiheadCrossAttention network for drug-target interaction prediction.

Accurate and effective drug-target interaction (DTI) prediction can greatly shorten the drug development lifecycle and reduce the cost of drug development. In the deep-learning-based paradigm for predicting DTI, robust drug and protein feature representations and their interaction features play a key role in improving the accuracy of DTI prediction. Additionally, the class imbalance problem and the overfitting problem in the drug-target dataset can also affect the prediction accuracy, and reducing the consumption of computational resources and speeding up the training process are also critical considerations. In this paper, we propose shared-weight-based MultiheadCrossAttention, a precise and concise attention mechanism that can establish the association between target and drug, making our models more accurate and faster. Then, we use the cross-attention mechanism to construct two models: MCANet and MCANet-B. In MCANet, the cross-attention mechanism is used to extract the interaction features between drugs and proteins for improving the feature representation ability of drugs and proteins, and the PolyLoss loss function is applied to alleviate the overfitting problem and the class imbalance problem in the drug-target dataset. In MCANet-B, the robustness of the model is improved by combining multiple MCANet models and prediction accuracy further increases. We train and evaluate our proposed methods on six public drug-target datasets and achieve state-of-the-art results. In comparison with other baselines, MCANet saves considerable computational resources while maintaining accuracy in the leading position; however, MCANet-B greatly improves prediction accuracy by combining multiple models while maintaining a balance between computational resource consumption and prediction accuracy.

Cross-task feature enhancement strategy in multi-task learning for harvesting Sichuan pepper

It is challenge to accurately perceive multi-objective functions in an unstructured, dynamic, and undefined agricultural environment. Owing to the neglect of multi-task relationships, some existing multi-task networks (MTN) for agricultural perception lack enough feature representation ability to perform robust prediction in an unstructured environment. To assist automatic harvesting of the Sichuan pepper, we develop a perception model and improve it by propose a cross-task feature enhancement (CTFE) strategy to exploit the multi-task relationships. CTFE comprises a feature harmonization module (FHM) and a series of asymmetrical feature sharing and fusing module (AFSFM). The FHM is devised to reconstruct the semantic feature from the object detection decoder. To this end, the semantic feature was aligned with the segmentation features in terms of scale and channel, which enabled the segmentation decoder to exploit semantic information from the detect decoder to filter background noise and activate responses of the shielded area. Subsequently, the AFSFM was embedded into two segmentation decoders to capture the complementary information used for segmentation boundaries alignment. The visualization results indicate that AFSFM can assist to spread the semantic information. Finally, the improved perception model was used to perform three tasks: foreground pepper detection, pepper segmentation, and stem segmentation. Experiments on the Sichuan pepper dataset demonstrate that the proposed CTFE strategy significantly improves the multi-task performance by transferring semantic features to suppress segmentation noise and enhance feature representation. Moreover, the proposed strategy yields competitive performance in each task and can assist generic multi-task learning methods to further enhance multi-task performance.

Hybrid unsupervised web-attack detection and classification – A deep learning approach

Web requests made by users of web applications are manipulated by hackers to gain control of web servers. Moreover, detecting web attacks has been increasingly important in the distribution of information over the last few decades. Also, several existing techniques had been performed on detecting vulnerable web attacks using machine learning and deep learning techniques. However, there is a lack in achieving attack detection ratio owing to the utilization of supervised and semi-supervised learning approaches. Thus to overcome the aforementioned issues, this research proposes a hybrid unsupervised detection model a deep learning-based anomaly-based web attack detection. Whereas, the encoded outputs of De-Noising Autoencoder (DAE), as well as Stacked Autoencoder (SAE), are integrated and given to the Generative adversarial network (GAN) as input to improve the feature representation ability to detect the web attacks. Consequently, for classifying the type of attacks, a novel DBM-Bi LSTM-based classification model has been introduced. Which incorporates DBM for binary classification and Bi-LSTM for multi-class classification to classify the various attacks. Finally, the performance of the classifier in terms of recall, precision, F1-Score, and accuracy are evaluated and compared. The proposed method achieved high accuracy of 98%.

Video Anomaly Detection Based on Attention Mechanism

Camera surveillance is widely used in residential areas, highways, schools and other public places. The monitoring and scanning of sudden abnormal events depend on humans. Human anomaly monitoring not only consumes a lot of manpower and time but also has a large error in anomaly detection. Video anomaly detection based on AE (Auto-Encoder) is currently the dominant research approach. The model has a highly symmetrical network structure in the encoding and decoding stages. The model is trained by learning standard video sequences, and the anomalous events are later determined in terms of reconstruction error and prediction error. However, in the case of limited computing power, the complex model will greatly reduce the detection efficiency, and unnecessary background information will seriously affect the detection accuracy of the model. This paper uses the AE loaded with dynamic prototype units as the basic model. We introduce an attention mechanism to improve the feature representation ability of the model. Deep separable convolution operation can effectively reduce the number of model parameters and complexity. Finally, we conducted experiments on three publicly available datasets of real scenarios (UCSD Ped1, UCSD Ped2 and CUHK Avenue). The experimental results show that compared with the baseline model, the accuracy of our model improved by 1.9%, 1.4% and 6.6%, respectively, across the three datasets. Compared with many popular models, the validity of our model in anomaly detection is verified.