A fabric defect detection model based on improved Swin Transformer with large kernel attention

Fabric inspection is critical in fabric manufacturing. Models of the R-CNN series perform extremely well in sophisticated fabric defects. However, the traditional CNNs are difficult to extract global features interrelationship, resulting in poor detection performance in large aspect ratio defects. In this paper, we based on the Cascade RCNN architecture that replaces the CNN backbones with Swin Transformer blocks for fabric defect detection. In addition, using the Large Kernel Attention in the detection head can use fewer computing resources to obtain a larger Receptive field. The experimental results show that the proposed SWLNet model’s mean average precision(mAP) is 64.60%, which improved accuracy by 2% compared to the current SOTA model IDPNet, but the detection efficiency is nearly double. By performing a comprehensive ablation study on ‘TianChi’ datasets, we demonstrate each module contributes to the overall improvement.