A Novel Fabric Defect Detection Network Based on Attention Mechanism and Multi-Task Fusion

Fabric is an important material, which is applied in the entire process of textile manufacturing, such as spinning, weaving, dyeing, printing, and finishing, and garments manufacturing. As defects on the surface of the fabric are inevitable in the process of fabric production, the defect detection of fabric is significant for fabric manufacture. The current CNN-based defect detection methods face several challenges when tackling the fabric defect with a tiny shape, the low grayscale difference with background, and ambiguous defect type. To deal with the problem, we proposed a novel fabric defect detection network – AMTFNet based on the attention mechanism and multi-task fusion module in this paper. On one hand, the attention mechanism module forced networks to pay attention to defects. On the other hand, the multi-task fusion module helps AMTFNet to further improve the classification effect using feature concatenated. The experimental result indicates that the precision-score, recall-score, and F1-score of AMTFNet reach 0.980, 0.994, and 0.987, respectively. The proposed method can be successfully applied in the detection of industrial fabric material.