Fabric Defect Detection Based On Nested Multi-Branch Network

In the process of textile production,defect detection plays a crucial role in ensuring product quality.Automated and intelligent defect detection methods can improve production efficiency,reduce human error rates,and lower production costs.Therefore,researching fabric defect detection techniques holds significant applications and economic value.With the development of the textile industry,fabric patterns have become increasingly complex,posing challenges for defect detection.This study aims to explore defect detection methods for complex patterned fabrics by utilizing techniques such as image enhancement,backbone network feature extraction,and feature fusion,with the goal of enhancing detection accuracy.The specific research contents are as follows:(1)Research on fabric defect detection method based on nested fusion of multi-scale information with dual-input architecture.To address the issues of misjudgment in complex texture regions and blurred edge detection in fabric defect detection,a dual-input network architecture called WNet is proposed.The network employs two backbone branches to extract multi-scale local and global features.Leveraging the global modeling capability of selfattention mechanism,it supplements global information in the deep convolutional network,reducing redundancy in texture features.To minimize the loss of local detail information in the deep network,a lightweight dual-branch pooling pyramid is introduced to incorporate shallowlevel multi-scale detail features into the deep modules.To integrate feature information at various scales,a multi-scale nested dual-branch module is proposed for adaptive fusion of multi-level receptive field features.Experiments show that the method achieves an average F1 value of 74.79%,an average recall of 78.68%,and an average precision of 71.30%.(2)Research on fabric defect detection method based on fusion of self-attention mechanism and recursive multi-level residual.To improve the accuracy of defect detection in complex pattern fabrics,a recursive multi-level residual hybrid network called U-SMR,incorporating Swin Transformer,is proposed.To address the issue of insufficient samples in periodic pattern datasets,a method for lattice disorder reconstruction based on adaptive segmentation algorithm is proposed.U-SMR adopts a nested U-shaped structure,enhancing the multi-scale global feature extraction capability of the backbone branch through the mixed connection of Swin Transformer Blocks.In the feature fusion stage,a recursive multi-level residual module is introduced,allowing flexible adjustment of module depth to accommodate different requirements.Experimental results demonstrate that U-SMR effectively handles complex texture fabric defects and achieves satisfactory performance.The average F1 value on the ZJU-Leaper dataset reached 75.33%,the average recall reached 78.54%,and the average precision reached 72.39%.The F1 value on the HKU-Fabric dataset reached 70.80%.(3)Research on fabric defect detection method based on dual-branch asymmetric convolutional attention.To address the issue of missed detection of small defects in complex pattern fabrics,a network called DBACA-Net,based on dual-branch asymmetric convolutional attention mechanism,is proposed.This study optimizes the backbone networks of WNet and U-SMR by incorporating Swin Transformer Blocks into the C2 channel connection of Res Net-18,reducing parameter count while preserving local and global semantic information.To improve the detection accuracy of small defects,a dual-branch convolutional attention pyramid module is proposed,leveraging multi-scale convolution and asymmetric convolutional attention mechanism to enhance the network’s sensitivity to stripe-like small defects.Experimental results demonstrate that DBACA-Net outperforms the previously proposed WNet and U-SMR methods significantly on the ZJU-Leaper dataset,achieving a substantial improvement in detection accuracy while reducing model computational cost.On the ZJULeaper dataset,DBACA-Net achieves an average F1 value of 75.66%,an average recall rate of79.12%,and an average precision rate of 72.51%.Compared with the previous WNet and USMR methods,DBACA-Net has improved in F1 value,completeness and accuracy.