Cloth Defect Detection Network Based on Feature Subtraction and Attention Mechanism

Recently, with the rapid development of deep learning, object detection has made significant progress. Though there are many excellent object detection algorithms, they still cannot achieve satisfactory results when applied to the detection of cloth defects with complex images directly. Due to the influence of complex images in the background, there exists a large number of missed detections and false detections, which also affects the location of defects. In this work, we make both normal cloth and defective cloth input into the feature extraction network. Then the respective features are obtained and subtracted, so as to remove the influence of background. The defects on the cloth are often too small and have large difference in shape. However, the propagation paths of previous feature pyramid networks are usually lateral and top-down, which means the information can only flow downwards and low-level information is not well utilized, resulting in low accuracy when detecting these defects. This paper improves the feature pyramid structure by fusing each layer in the feature pyramid with its adjacent layers and introducing an attention mechanism to allow the network to learn information more effectively, thus improving the performance. We also modified the loss function of the final classification-regression network to improve the influence of easily separable targets on model parameters. After extensive experiments, our Feature Subtraction and Attention Cascade RCNN (FSA-Cascade RCNN) network achieves 57.18% mAP and 92.12% Acc, which is superior than existing algorithms.