| The existence of fabric surface defects affects the quality of textiles,makes the price of products drop significantly,and seriously affects the efficiency of enterprises.The traditional method uses the way of manual visual inspection to detect fabric defects,which requires high experience,low work efficiency and missing inspection.With the increasing diversity of fabric defects,it is an urgent problem to design an efficient and high-precision fabric defect detection system.Deep learning method has become the mainstream of research because of its powerful ability of automatic feature extraction.It brings a new method for defect detection and has been widely used in industrial product defect detection.This paper discusses the application of deep learning in fabric defect detection with complex texture,and proposes two fabric defect detection methods based on convolution neural network.(1)This paper designs a fabric defect classification algorithm based on spatial attention hole convolution feature fusion model.Different void rates are selected for superposition to effectively maintain the integrity of feature map information,and the spatial attention module is integrated into convolutional neural network to learn the context dependence of defects,extract the key location information,enhance the differential expression of features,and further improve the classification accuracy;the classification network can only obtain the classification information of defects,and additional location information is needed to obtain the location information The improved class activation mapping method is used to obtain the defect classification information and location information in the case of only image level annotation.On the basis of data enhancement and transfer learning method,the fabric surface defect image is recognized and detected.The experimental results show that the algorithm can effectively improve the accuracy of fabric defect classification,and can obtain the defect location information without manual location labeling.(2)In this paper,a fabric defect segmentation algorithm based on attention guided feature fusion is designed.In the coding stage,multi-layer convolution pooling is used to obtain the high-level semantic information of the image,and a self attention pyramid pooling module is designed at the end of the encoder,which integrates local features and global features while building rich context information to further enhance the feature representation.In the decoder stage,a new path from top to bottom is established,through which the initial multi-scale information of each layer features can be obtained.In the cross layer structure,a special feature fusion network is designed.The large convolution kernel symmetric convolution is used to improve the spatial attention mechanism to eliminate the interference of shallow feature background texture information and highlight the defect information.The residual structure and channel attention are used to act on the feature fusion process,and the relationship between channels is modeled to guide the network to select useful features To achieve better information fusion of deep features and shallow features.The experimental results show that,compared with the traditional segmentation algorithm,the proposed method improves the pixel accuracy,the average pixel accuracy and the average intersection and union ratio,which shows the effectiveness of the algorithm. |
