Features Of Different Dimensions Research Articles

GS-DeepLabV3+: A Mountain Tea Disease Segmentation Network Based on Improved Shuffle Attention and Gated Multidimensional Feature Extraction

The prevention and control of Camellia Oleifera diseases can produce significantly comprehensive economic benefits. And precisely identifying the types of diseases on oil tea leaves can enhance the effectiveness of pest and disease control. Currently, the most effective deep learning methods for segmentation are suffering from the challenges such as indistinct disease spots on the actual oil tea leaves that are difficult to differentiate from the background, the high similarities in characteristics between different diseases, and the small target area of disease regions that are easily ignored. In this paper, GS-DeepLabV3+ network is proposed, which is an improvement upon DeepLabV3+ and aims to effectively enhance disease identification and segmentation rates. Initially, a gated pyramid feature fusion structure is proposed, utilizing a special module to merge features of different dimensions to enhance differentiation of similar features. Secondly, an improved grouped attention fusion mechanism is introduced, which allocates channel and spatial weights to different feature groups, enhancing the extraction of detailed features. Lastly, we substituted the backbone network of DeepLabV3+ with the more lightweight MobileNetV2, which significantly reduces the number of parameters while more stably extracting multidimensional features. In comparison with the existing segmentation networks on a self-created oil tea leaf disease data set, GS-DeepLabV3+ achieved a mIoU of 87.77% and an average pixel accuracy of 94.55%. Compared to the base DeepLabV3+ network, there was a 4.14% improvement in mIoU, a 3.58% increase in mPA, and a 1.4% rise in precision. The improved attention mechanism contributed to a 1.98% increase in mIoU, and the gated multidimensional fusion mechanism resulted in a 1.03% improvement. These enhancements can provide technical references for crop protection and pest disease segmentation.

Gearbox fault diagnosis method based on lightweight channel attention mechanism and transfer learning

In practical engineering, the working conditions of gearbox are complex and variable. In varying working conditions, the performance of intelligent fault diagnosis model is degraded because of limited valid samples and large data distribution differences of gearbox signals. Based on these issues, this research proposes a gearbox fault diagnosis method integrated with lightweight channel attention mechanism, and further realizes the cross-component transfer learning. First, time–frequency distribution of original signals is obtained by wavelet transform. It could intuitively reflect local characteristics of signals. Secondly, based on a local cross-channel interaction strategy, a lightweight efficient channel attention mechanism (LECA) is designed. The kernel size of 1D convolution is affected by channel number and coefficients. Multi-scale feature input is used to retain more detailed features of different dimensions. A lightweight convolutional neural network is constructed. Finally, a transfer learning method is applied to freeze lower structures of the network and fine-tune higher structures of the model using small samples. Through experimental verification, the proposed model could effectively utilize samples. The application of transfer learning could realize accurate and fast fault classification of small samples, and achieve good gearbox fault diagnosis effect under varying working conditions and cross-component conditions.

Feature extraction and representation learning of 3D point cloud data

Three-dimensional point cloud data serves as a critical source of information in various real-world application domains, such as computer vision, robotics, geographic information systems, and medical image processing. Due to the discrete and unordered nature of point clouds, applying 2D image feature extractors directly to the extraction of 3D point cloud features is challenging. Therefore, we propose a novel variational feature component extraction method called PointFEA. This paper aims to research and propose a series of methods to enhance the feature extraction and representation learning of 3D point cloud data. Firstly, in terms of feature extraction, local neighborhood encoding is combined with the local latent representation of point clouds to obtain more correlated point cloud features. Secondly, in the domain of point cloud representation learning, the multi-scale representation learning method maps point cloud data into a high-dimensional space to better capture critical features and adapt to different granularities of point cloud data. Lastly, features of different dimensions are input into a cross-fusion transformer to obtain local attention coefficients. We validate our methods on commonly used point cloud datasets, and the experiments demonstrate the effectiveness of our approach, achieving accuracies of 94.8% on ModelNet40 and 89.1% on ScanObjectNN.

Meta-adaptive-weighting-based bilateral multi-dimensional refined space feature attention network for imbalanced breast cancer histopathological image classification

Breast cancer histopathological image automatic classification can reduce pathologists workload and provide accurate diagnosis. However, one challenge is that empirical datasets are usually imbalanced, resulting in poorer classification quality compared with conventional methods based on balanced datasets. The recently proposed bilateral branch network (BBN) tackles this problem through considering both representation and classifier learning to improve classification performance. We firstly apply bilateral sampling strategy to imbalanced breast cancer histopathological image classification and propose a meta-adaptive-weighting-based bilateral multi-dimensional refined space feature attention network (MAW-BMRSFAN). The model is composed of BMRSFAN and MAWN. Specifically, the refined space feature attention module (RSFAM) is based on convolutional long short-term memories (ConvLSTMs). It is designed to extract refined spatial features of different dimensions for image classification and is inserted into different layers of classification model. Meanwhile, the MAWN is proposed to model the mapping from a balanced meta-dataset to imbalanced dataset. It finds suitable weighting parameter for BMRSFAN more flexibly through adaptively learning from a small amount of balanced dataset directly. The experiments show that MAW-BMRSFAN performs better than previous methods. The recognition accuracy of MAW-BMRSFAN under four different magnifications still is higher than 80% even when unbalance factor is 16, indicating that MAW-BMRSFAN can make ideal performance under extreme imbalanced conditions.

DA-FER: Domain Adaptive Facial Expression Recognition

Facial expression recognition (FER) is an important field in computer vision with many practical applications. However, one of the challenges in FER is dealing with small sample data, where the number of samples available for training machine learning algorithms is limited. To address this issue, a domain adaptive learning strategy is proposed in this paper. The approach uses a public dataset with sufficient samples as the source domain and a small sample dataset as the target domain. Furthermore, the maximum mean discrepancy with kernel mean embedding is utilized to reduce the disparity between the source and target domain data samples, thereby enhancing expression recognition accuracy. The proposed Domain Adaptive Facial Expression Recognition (DA-FER) method integrates the SSPP module and Slice module to fuse expression features of different dimensions. Moreover, this method retains the regions of interest of the five senses to accomplish more discriminative feature extraction and improve the transfer learning capability of the network. Experimental results indicate that the proposed method can effectively enhance the performance of expression recognition. Specifically, when the self-collected Selfie-Expression dataset is used as the target domain, and the public datasets RAF-DB and Fer2013 are used as the source domain, the performance of expression recognition is improved to varying degrees, which demonstrates the effectiveness of this domain adaptive method.

Semantic-aware knowledge distillation with parameter-free feature uniformization

Knowledge distillation aims to distill knowledge from teacher networks to train student networks. Distilling intermediate features has attracted much attention in recent years as it can be flexibly applied in various fields such as image classification, object detection and semantic segmentation. A critical obstacle of feature-based knowledge distillation is the dimension gap between the intermediate features of teacher and student, and plenty of methods have been proposed to resolve this problem. However, these works usually implement feature uniformization in an unsupervised way, lacking guidance to help the student network learn meaningful mapping functions in the uniformization process. Moreover, the dimension uniformization process of the student and teacher network is usually not equivalent as the mapping functions are different. To this end, some factors of the feature are discarded during parametric feature alignment, or some factors are blended in some non-parametric operations. In this paper, we propose a novel semantic-aware knowledge distillation scheme to solve these problems. We build a standalone feature-based classification branch to extract semantic-aware knowledge for better guiding the learning process of the student network. In addition, to avoid the inequivalence of feature uniformization between teacher and student, we design a novel parameter-free self-attention operation that can convert features of different dimensions into vectors of the same length. Experimental results show that the proposed knowledge distillation scheme outperforms existing feature-based distillation methods on the widely used CIFAR-100 and CINIC-10 datasets.

Learning From Heterogeneous Data Based on Social Interactions Over Graphs

This work proposes a decentralized architecture, where individual agents aim at solving a classification problem while observing streaming features of different dimensions and arising from possibly different distributions. In the context of social learning, several useful strategies have been developed, which solve decision making problems through local cooperation across distributed agents and allow them to learn from streaming data. However, traditional social learning strategies rely on the fundamental assumption that each agent has significant prior knowledge of the underlying distribution of the observations. In this work we overcome this issue by introducing a machine learning framework that exploits social interactions over a graph, leading to a fully data-driven solution to the distributed classification problem. In the proposed social machine learning (SML) strategy, two phases are present: in the training phase, classifiers are independently trained to generate a belief over a set of hypotheses using a finite number of training samples; in the prediction phase, classifiers evaluate streaming unlabeled observations and share their instantaneous beliefs with neighboring classifiers. We show that the SML strategy enables the agents to learn consistently under this highly-heterogeneous setting and allows the network to continue learning even during the prediction phase when it is deciding on unlabeled samples. The prediction decisions are used to continually improve performance thereafter in a manner that is markedly different from most existing static classification schemes where, following training, the decisions on unlabeled data are not re-used to improve future performance.

A Multi-Scale Temporal Convolutional Network with Attention Mechanism for Force Level Classification during Motor Imagery of Unilateral Upper-Limb Movements.

In motor imagery (MI) brain-computer interface (BCI) research, some researchers have designed MI paradigms of force under a unilateral upper-limb static state. It is difficult to apply these paradigms to the dynamic force interaction process between the robot and the patient in a brain-controlled rehabilitation robot system, which needs to induce thinking states of the patient's demand for assistance. Therefore, in our research, according to the movement of wiping the table in human daily life, we designed a three-level-force MI paradigm under a unilateral upper-limb dynamic state. Based on the event-related de-synchronization (ERD) feature analysis of the electroencephalography (EEG) signals generated by the brain's force change motor imagination, we proposed a multi-scale temporal convolutional network with attention mechanism (MSTCN-AM) algorithm to recognize ERD features of MI-EEG signals. Aiming at the slight feature differences of single-trial MI-EEG signals among different levels of force, the MSTCN module was designed to extract fine-grained features of different dimensions in the time-frequency domain. The spatial convolution module was then used to learn the area differences of space domain features. Finally, the attention mechanism dynamically weighted the time-frequency-space domain features to improve the algorithm's sensitivity. The results showed that the accuracy of the algorithm was 86.4 ± 14.0% for the three-level-force MI-EEG data collected experimentally. Compared with the baseline algorithms (OVR-CSP+SVM (77.6 ± 14.5%), Deep ConvNet (75.3 ± 12.3%), Shallow ConvNet (77.6 ± 11.8%), EEGNet (82.3 ± 13.8%), and SCNN-BiLSTM (69.1 ± 16.8%)), our algorithm had higher classification accuracy with significant differences and better fitting performance.

Music Emotion Recognition Based on a Neural Network with an Inception-GRU Residual Structure

As a key field in music information retrieval, music emotion recognition is indeed a challenging task. To enhance the accuracy of music emotion classification and recognition, this paper uses the idea of inception structure to use different receptive fields to extract features of different dimensions and perform compression, expansion, and recompression operations to mine more effective features and connect the timing signals in the residual network to the GRU module to extract timing features. A one-dimensional (1D) residual Convolutional Neural Network (CNN) with an improved Inception module and Gate Recurrent Unit (GRU) was presented and tested on the Soundtrack dataset. Fast Fourier Transform (FFT) was used to process the samples experimentally and determine their spectral characteristics. Compared with the shallow learning methods such as support vector machine and random forest and the deep learning method based on Visual Geometry Group (VGG) CNN proposed by Sarkar et al., the proposed deep learning method of the 1D CNN with the Inception-GRU residual structure demonstrated better performance in music emotion recognition and classification tasks, achieving an accuracy of 84%.

TREPH: A Plug-In Topological Layer for Graph Neural Networks

Topological Data Analysis (TDA) is an approach to analyzing the shape of data using techniques from algebraic topology. The staple of TDA is Persistent Homology (PH). Recent years have seen a trend of combining PH and Graph Neural Networks (GNNs) in an end-to-end manner to capture topological features from graph data. Though effective, these methods are limited by the shortcomings of PH: incomplete topological information and irregular output format. Extended Persistent Homology (EPH), as a variant of PH, addresses these problems elegantly. In this paper, we propose a plug-in topological layer for GNNs, termed Topological Representation with Extended Persistent Homology (TREPH). Taking advantage of the uniformity of EPH, a novel aggregation mechanism is designed to collate topological features of different dimensions to the local positions determining their living processes. The proposed layer is provably differentiable and more expressive than PH-based representations, which in turn is strictly stronger than message-passing GNNs in expressive power. Experiments on real-world graph classification tasks demonstrate the competitiveness of TREPH compared with the state-of-the-art approaches.

Spatial-Temporal Feature with Dual-Attention Mechanism for Encrypted Malicious Traffic Detection

While encryption ensures the confidentiality and integrity of user data, more and more attackers try to hide attack behaviours through encryption, which brings new challenges to malicious traffic identification. How to effectively detect encrypted malicious traffic without decrypting traffic and protecting user privacy has become an urgent problem to be solved. Most of the current research only uses a single CNN, RNN, and SAE network to detect encrypted malicious traffic, which does not consider the forward and backward correlation between data packets, so it is difficult to effectively identify malicious features in encrypted traffic. This study proposes an approach that combines spatial-temporal feature with dual-attention mechanism, which is called TLARNN. Specifically, first we use 1D-CNN and BiGRU to extract spatial features in encrypted traffic packets and temporal features between encrypted streams, respectively, which enriches the features of different dimensions, and then, the soft attention mechanism is focused on the encrypted data packets to extract features. Ultimately, the second layer of the soft attention mechanism is used for aggregating malicious features. Several comparative experiments are designed to prove the effectiveness of the proposed scheme. The experimental results demonstrate that the proposed scheme has a significant performance improvement compared to existing ones.

Image Inpainting Based on Structural Constraint and Multi-Scale Feature Fusion

When repairing masked images based on deep learning, there is usually insufficient representation of multi-level information and inadequate utilization of long distance features. To solve the problems, this paper proposes a second-order generative image inpainting model based on Structural Constraints and Multi-scale Feature Fusion (SCMFF). The SCMFF model consists of two parts: edge repair network and image inpainting network. The edge repair network combines the auto-encoder with the Dilated Residual Feature Pyramid Fusion (DRFPF) module, which improves the representation of multi-level semantic information and structural details of images, thus achieves better edge repair. Then, the image inpainting network embeds the Dilated Multi-scale Attention Fusion (DMAF) module in the auto-encoder for texture synthesis with the real edge as the prior condition, and achieves fine-grained inpainting under the edge constraint by aggregating the long-distance features of different dimensions. Finally, the edge repair results are used to replace the real edge, and the two networks are fused and trained to achieve end-to-end repair from the masked image to the complete image. The model is compared with the advanced methods on datasets including Celeba, Facade and Places2. The quantitative results show that the four metrics of LPIPS, MAE, PSNR and SSIM are improved by 0.0124-0.0211, 3.787-6.829, 2.934dB-5.730dB and 0.034-0.132, respectively. The qualitative results show that the edge distribution in the center of the hole reconstructed by the SCMFF model is more uniform, and the texture synthesis effect is more in line with human visual perception.

FCDM: An Improved Forest Fire Classification and Detection Model Based on YOLOv5

Intense, large-scale forest fires are damaging and very challenging to control. Locations, where various types of fire behavior occur, vary depending on environmental factors. According to the burning site of forest fires and the degree of damage, this paper considers the classification and identification of surface fires and canopy fires. Deep learning-based forest fire detection uses convolutional neural networks to automatically extract multidimensional features of forest fire images with high detection accuracy. To accurately identify different forest fire types in complex backgrounds, an improved forest fire classification and detection model (FCDM) based on YOLOv5 is presented in this paper, which uses image-based data. By changing the YOLOv5 bounding box loss function to SIoU Loss and introducing directionality in the cost of the loss function to achieve faster convergence, the training and inference of the detection algorithm are greatly improved. The Convolutional Block Attention Module (CBAM) is introduced in the network to fuse channel attention and spatial attention to improve the classification recognition accuracy. The Path Aggregation Network (PANet) layer in the YOLOv5 algorithm is improved into a weighted Bi-directional Feature Pyramid Network (BiFPN) to fuse and filter forest fire features of different dimensions to improve the detection of different types of forest fires. The experimental results show that this improved forest fire classification and identification model outperforms the YOLOv5 algorithm in both detection performances. The mAP@0.5 of fire detection, surface fire detection, and canopy fire detection was improved by 3.9%, 4.0%, and 3.8%, respectively. Among them, the mAP@0.5 of surface fire reached 83.1%, and the canopy fire detection reached 90.6%. This indicates that the performance of our proposed improved model has been effectively improved and has some application prospects in forest fire classification and recognition.

A Novel Threat Intelligence Information Extraction System Combining Multiple Models

The increasing number of cyberattacks has made the cybersecurity situation more serious. Thus, it is urgent to use cyber threat intelligence to deal with the complex and changing cyber environment. However, cyber threat intelligence usually exists in an unstructured form, and a huge amount of data poses a great challenge to security analysts. To this end, this paper proposes a novel threat intelligence information extraction system combining multiple models, which contains four key steps: entity extraction, coreference resolution, relation extraction, and knowledge graph construction. In the entity extraction task, a multihead self-attention mechanism is adopted to extract the dependency relationships between words. In the coreference resolution task, contextual information and mention embedding are fused to improve the mention representation. Meanwhile, features of different dimensions are extracted using a convolutional neural network. In the relation extraction task, additional features such as part of speech, mention width, entity type, and distance of entity pairs are incorporated to improve the embedding representation. Finally, a knowledge graph is constructed to explicitly present entities and their relationships. Experimental results indicate that compared with the baseline model, the F1 score of our model is improved by at least 8.87, 9.82, and 10.56 on entity extraction, coreference resolution, and relation extraction, respectively. The knowledge graph in Neo4j demonstrates the effectiveness of our system.

Multi-dimensional shared representation learning with graph fusion network for Session-based Recommendation

The Session-based Recommendation (SBR) system aims to forecast anonymous users’ short-term decisions. Many prior research have demonstrated that using Graph Neural Networks (GNN) and Recurrent Neural Networks (RNN) to solve SBR tasks can lead to excellent results. However, the existing SBR models only use single feature extraction method to represent an item in the recommendation process. These models either use the GNN methods which ignore the sequence location features in the training process or use RNN methods which cannot obtain the weight features of every item in the sequence. All of them fail to extract features of different dimensions at the same time, which results in low-quality item representations. To tackle the aforementioned issue, this research introduces a novel method called Multi-dimensional Shared Representation Learning (MSR). (i) To get multi-dimensional features, we use the multi-dimensional feature extraction in a double-flow way based on transformer layers and graph attention network (GAT) layers. (ii) Through the MSR module, we train the joint representation of items in the session by using multiple GNNs, then merge the multi-dimensional features by using an attention mechanism. A variety of experiments have been carried out and the simulation results demonstrate the designed strategy largely surpasses the state-of-the-art proposal in recommendation accuracy.

A Network Model for Detecting Marine Floating Weak Targets Based on Multimodal Data Fusion of Radar Echoes

Due to the interaction between floating weak targets and sea clutter in complex marine environments, it is necessary to distinguish targets and sea clutter from different dimensions by designing universal deep learning models. Therefore, in this paper, we introduce the concept of multimodal data fusion from the field of artificial intelligence (AI) to the marine target detection task. Using deep learning methods, a target detection network model based on the multimodal data fusion of radar echoes is proposed. In the paper, according to the characteristics of different modalities data, the temporal LeNet (T-LeNet) network module and time-frequency feature extraction network module are constructed to extract the time domain features, frequency domain features, and time-frequency features from radar sea surface echo signals. To avoid the impact of redundant features between different modalities data on detection performance, a Self-Attention mechanism is introduced to fuse and optimize the features of different dimensions. The experimental results based on the publicly available IPIX radar and CSIR datasets show that the multimodal data fusion of radar echoes can effectively improve the detection performance of marine floating weak targets. The proposed model has a target detection probability of 0.97 when the false alarm probability is 10-3 under the lower signal-to-clutter ratio (SCR) sea state. Compared with the feature-based detector and the detection model based on single-modality data, the new model proposed by us has stronger detection performance and universality under various marine detection environments. Moreover, the transfer learning method is used to train the new model in this paper, which effectively reduces the model training time. This provides the possibility of applying deep learning methods to real-time target detection at sea.

Development of Weed Detection Method in Soybean Fields Utilizing Improved DeepLabv3+ Platform

Accurately identifying weeds in crop fields is key to achieving selective herbicide spraying. Weed identification is made difficult by the dense distribution of weeds and crops, which makes boundary segmentation at the overlap inaccurate, and thus pixels cannot be correctly classified. To solve this problem, this study proposes a soybean field weed recognition model based on an improved DeepLabv3+ model, which uses a Swin transformer as the feature extraction backbone to enhance the model’s utilization of global information relationships, fuses feature maps of different sizes in the decoding section to enhance the utilization of features of different dimensions, and adds a convolution block attention module (CBAM) after each feature fusion to enhance the model’s utilization of focused information in the feature maps, resulting in a new weed recognition model, Swin-DeepLab. Using this model to identify a dataset containing a large number of densely distributed weedy soybean seedlings, the average intersection ratio reached 91.53%, the accuracy improved by 2.94% compared with that before the improvement with only a 48 ms increase in recognition time, and the accuracy was superior to those of other classical semantic segmentation models. The results showed that the Swin-DeepLab network proposed in this paper can successfully solve the problems of incorrect boundary contour recognition when weeds are densely distributed with crops and incorrect classification when recognition targets overlap, providing a direction for the further application of transformers in weed recognition.

Click-Through Rate Prediction Algorithm Based on Modeling of Implicit High-Order Feature Importance

<p>Click-through rate (CTR) prediction plays a central role in online advertising and recommendation systems. In recent years, with the successful application of deep neural networks (DNNs) in many fields, researchers have integrated deep learning into CTR prediction algorithms to model implicit high-order features. However, most of these existing methods unify the weights of implicit higher-order features to predict user behaviors. The importance of such features of different dimensions for predicting user click behaviors are different. Base on this, we propose a prediction method that dynamically learns the importance of implicit high-order features. Specifically, we integrate the output features of deep and shallow components, and adaptively learn the weights of implicit high-order features from among all features through the designed attention network, which effectively capturing the deep interests of users. In addition, this framework has strong versatility and can be combined with shallow models such as Logistic Regression (LR) and Factorization Machines (FMs) to form different models and achieve optimal performance. The extended experiment is conducted on two large-scale datasets, AVAZU and SafeDrive, and the experimental results show that the performance of the proposed model is superior to that of existing baseline models.</p> <p> </p>

Improved inter-residue contact prediction via a hybrid generative model and dynamic loss function

Protein contact maps represent spatial pairwise inter-residue interactions, providing a protein's translationally and rotationally invariant topological representation. Accurate contact map prediction has been a critical driving force for improving protein structure determination. Contact maps can also be used as a stand-alone tool for varied applications such as prediction of protein–protein interactions, structure-aware thermal stability or physicochemical properties. We develop a novel hybrid contact map prediction model, CGAN-Cmap, that uses a generative adversarial neural network embedded with a series of modified squeeze and excitation residual networks. To exploit features of different dimensions, we introduce two parallel modules. This architecture improves the prediction by increasing receptive fields, surpassing redundant features and encouraging more meaningful ones from 1D and 2D inputs. We also introduce a new custom dynamic binary cross-entropy loss function to address the input imbalance problem for highly sparse long-range contacts in proteins with insufficient homologs. We evaluate the model’s performance on CASP 11, 12, 13, 14, and CAMEO test sets. CGAN-Cmap outperforms state-of-the-art models, improving precision of medium and long-range contacts by at least 3.5%. As a direct assessment between our model and AlphaFold2, the leading available protein structure prediction model, we compare extracted contact maps from AlphaFold2 and predicted contact maps from CGAN-Cmap. The results show that CGAN-Cmap has a mean precision higher by 1% compared to AlphaFold2 for most ranges of contacts. These results demonstrate an efficient approach for highly accurate contact map prediction toward accurate characterization of protein structure, properties and functions from sequence.

Application of YOLOv5 Based on Attention Mechanism and Receptive Field in Identifying Defects of Thangka Images

Aiming at the problems of target detection network in the defect detection field of thangka images with complex background colors, such as poor small target detection effect, insufficient feature information extraction, prone to error detection and leak detection, and low accuracy of defect detection, this paper proposed the YOLOv5 defect detection algorithm combining attention mechanism and receptive field. First of all, the Backbone network is used for feature extraction, integrating attention mechanism to represent different features, so that the network can fully extract the texture and semantic features of the defect area, and the extracted features are weighted and fused to reduce information loss. Secondly, a weighted fusion of features of different dimensions is transferred by the Neck network, and the combination of FPN and PAN is used to realize the fusion of semantic features and texture features of different layers and to locate the defect target more accurately. Finally, while replacing the GIoU loss function with CIoU, the receptive field is added to the network, so that the algorithm uses a four-channel detection mechanism to expand the detection range of receptive fields, and fuses semantic information between different network layers, so as to achieve fast location and more refined processing of small targets. The experimental results show that compared with the original YOLOv5 network, the detection accuracy of YOLOV5-scSE and YOLOV5-CA networks proposed in this paper has improved by 9.96 percentage points and 12.22 percentage points respectively, and the verification index has been significantly improved. It can quickly and more accurately identify and locate the location of the defect area and has a stronger ability to generalize the defect category, which greatly improves the accuracy of thangka image defect detection.