Multi-scale defect detection for plaid fabrics using scale sequence feature fusion and triple encoding

Fabric defect detection is always challenging. The appearance of textile defects affects the use value of fabrics and can cause significant economic losses for textile factories. Many methods exist for fabric defect detection, but they often exhibit low precision in localizing defects, especially small ones, or when the defect scale varies significantly. Therefore, to solve the above problems, this paper proposes an improved YOLOv8 fabric defect detection algorithm, which is a recognition method for detecting different textile background defects. First, the algorithm incorporates the slim-neck architecture to reduce complexity while preserving accuracy. Additionally, it integrates the scale sequence feature fusion (SSFF) module and the triple feature encoding (TFE) module, enabling more precise feature extraction from intricate textile patterns. A novel Shape-IoU loss function is employed to refine bounding box regression, and the substitution of the traditional decoder with the DETR decoder further improves detection capabilities. The algorithm was evaluated on a dataset containing six defect types, with data augmentation and preprocessing. It achieved a mean average precision (mAP) of 0.954. Compared to models like YOLOv5s, YOLOv7, YOLOv7-tiny, RT-DETR, and the original YOLOv8, the proposed algorithm showed mAP improvements of 3.50%, 5.80%, 9.50%, 1.30%, and 2.60%, respectively. Additionally, it demonstrated 88.7% precision, 98% recall, faster convergence, and reduced error. To further demonstrate its generalizability, the algorithm was evaluated on the Alibaba Cloud Tianchi fabric defect dataset, achieving the mAP of 78.63%. The proposed algorithm outperforms state-of-the-art methods, better meeting the demands of modern textile factories. Our source code is available at: https://github.com/dodolearning/fabric_yolo/tree/master.