Research On Fabric Defect Based On Swin Transformer And NAS-FPN

China is a major exporter of textile products,and fabric defect detection is an important aspect of textile quality control.Effective automated fabric defect detection methods can improve product quality and enhance competitiveness of enterprises.Target detection-based fabric defect methods are a popular research direction.However,due to the complexity of fabric textures and the diversity of defects,conventional target detection algorithms perform poorly in defect detection,with low recognition accuracy.Additionally,the lack of large-scale defect image support makes it difficult to train a universal defect detection model.To address these two problems,this study primarily focuses on the following research work:Defect datasets suffer from issues such as limited quantity,class imbalance,and difficulty in collection.To solve this problem,this study proposes three data augmentation schemes specifically for defects.Firstly,techniques like rotation,flipping,translation,and scaling are used to increase the quantity of images.Next,a multi-level generative adversarial network with attention mechanism is constructed to generate specific defects by inputting noise.The pix2 pix algorithm is employed to better integrate defects into the background,enhancing the quality of the images.To address the issue of two-stage generative adversarial networks failing to generate small defects,a copy-paste method is adopted to improve the detection capability of small defects and better identify defects.To address the problem of poor performance in general object detection,a fabric defect detection network based on the Swin Transformer is proposed on top of the two-stage object detection network.The attention-based architecture serves as the backbone for feature extraction,allowing better focus on the relationship between defects and the background.Additionally,a neural architecture search is introduced to automatically search for more effective feature fusion connection structures.The design of anchors uses K-Means clustering to calculate the optimal anchor aspect ratio to adapt to extreme aspect ratios of defects.A cascade detection head with IOU thresholds of 0.5,0.6,and 0.7 is used to improve detection accuracy.The Soft-NMS method is introduced for post-processing to further enhance detection accuracy.Comparative and ablation experiments were conducted on a publicly available dataset with 6,899 images and 20 defect classes.The experimental results demonstrate that compared to the baseline model Cascade RCNN,the proposed method achieved an average detection accuracy of 0.575,a 31.8% improvement over the baseline model’s average detection accuracy of 0.436.