YOLOv4-DCN-based fabric defect detection algorithm

Abstract

Defect detection in fabric production is the key to ensuring the quality of fabric. At present, the manual detection methodology employed in the textile field has many shortcomings. Therefore, researchers began to introduce machine learning, deep learning, and other methods to achieve automatic fabric defect detection. However, existing object detection algorithms still have some shortcomings because the wide variety of fabric defects, unbalanced categories, different scales, a large proportion of small object defects, and extreme aspect ratios. In response to these problems, this paper proposes an improved YOLOv4 algorithm. Deformable convolution is added to the YOLOv4 backbone network, which strengthens the network’s ability to describe geometric variations. At the same time, various data augmentation methods, which greatly enriches the dataset, are performed on the original dataset. K-means clustering is also performed on the dataset to obtain more suitable anchors. Finally, the mAP grew by 11.6%, reaching 81.26%, the precision rate grew by 3.89%, and the recall rate grew by 17.1% when compared to the original YOLOv4 network. However, the detection speed decreased. The results show that all indexes of YOLOv4-DCN are improved and can be more effectively to detect fabric defects.