Fabric defect detection based on YOLOv8-TR

Abstract

Fabric defects are a key factor affecting fabric production quality and fabric grade. In recent years, deep learning-based fabric defect detection algorithms have been proposed, but these methods cannot be adapted to complex color-patterned fabrics, and suffer from leakage and false detection problems. Therefore, this article proposes a new algorithm based on the Transformer architecture, called YOLOv8-Transformer (YOLOv8-TR), which aims to achieve fast and accurate defect detection for color-patterned fabrics. The algorithm first utilizes the backbone network based on the Transformer architecture to obtain the global perceptual field to enhance the feature extraction capability. Second, a migration learning operation is performed when training the improved network after replacing the backbone network in order to avoid the problems of insufficient parameter optimization and poor model generalization when training the new network architecture and to improve the detection accuracy. Finally, the intersection union loss function as a localization loss is optimized to improve the localization accuracy of predicted frames. The proposed algorithm is compared with the existing methods on Alibaba Cloud Tianchi Fabric dataset, and YOLOv8-TR achieves a mean Average Precision of 81.5%, which is 5.2% higher than YOLOv8, with a detection speed of 135 frames per second. The experimental results verify the effectiveness and superiority of the method. A solution for automatic detection of fabric defects in the textile manufacturing industry is provided.