SRSE-YOLO: An improved fabric defect detection method for YOLOv8

Abstract

Aiming at the existing problems of fabric defect detection, the improved YOLOv8 fabric defect detection method SRSE-YOLO is proposed, which effectively improves the model’s ability to extract and utilize complex texture features while maintaining low computational cost, and is more suited to fabric defect detection tasks. To further lighten the model and reduce the computational complexity, this paper introduces SCConv to improve the partial convolution module in the backbone and proposes a new spatial pyramid pooling fusion module RL-SPPF combined with LSKA to simplify the model structure. In the Neck, SCE-C2f combined with an edge detection module is proposed, which addresses for the deficiency of traditional C2f module in dealing with small texture and edge features. In the Detection Head, the EMA attention mechanism is introduced to dynamically adjust the weights of each channel and improve the perception ability of the model to the defect features. Furthermore, Inner-DIoU is used to enhance the loss function, and a 160 × 160 detection head is added, which improves the detection ability of the model for tiny defects and accelerates the convergence of the model. The experimental results show that compared with the benchmark model YOLOv8, the precision (P), recall (R), mAP@0.5, and mAP@0.5:0.95 indicators of SRSE-YOLO on the Aliyun Tianchi dataset are increased by 5.5%, 4.0%, 3.71%, and 25.68%, respectively. SRSE-YOLO maintains the same level as the benchmark model in terms of parameter count and computational load, realizes the balance between detection performance and calculation resources, effectively improves the detection performance of fabric defects, and provides a new solution for fabric quality control.