A texture-aware one-stage fabric defect detection network with adaptive feature fusion and multi-task training

Fabric defect detection is an indispensable process to guarantee product quality in industrial production. With the proposal of industry 4.0, manufacturing enterprises have been endeavoring to develop automatic fabric defect detection systems to overcome the shortcomings of traditional manual defect detection, and many researchers have been devoting themselves to designing fabric defect detection methods with high effectiveness and strong anti-noise capacity. In the paper, we notice that the perception of the normal fabric textures is beneficial to detecting fabric defects, which is ignored in previous researches for classic one-stage detection network. Based on this finding, we propose a texture-aware one-stage fabric defect detection network (TADet). First, an adaptive feature fusion module is designed to merge and enhance multi-resolution features, where the attention mechanism is employed to make the feature fusion process adaptive to the input fabric images with different textures. Second, a multi-task defect detection head is proposed for forcing the network to distinguish the defect regions and texture regions while classifying and locating defects simultaneously in the training phase. With the defect-texture distinguishing tasks, the network is able to consider the fabric textures explicitly when detecting defects. The experimental results show that the proposed TADet outperforms other two-stage models by 3.7% and one-stage detection models by 9.3% on mAP. Besides, further experiments verify the high efficiency and strong noise-robustness of the proposed TADet, which shows its potential for industrial applications.