Improved YOLOv5s Fabric Defect Detection Algorithm Based on the Fusion of Global Attention Mechanism and Decoupling Head

Abstract

During fabric production, machine malfunctions and yarn breakages can lead to defects, affecting product quality. Thus, fabric defect detection is crucial for quality control. Existing detection systems struggle with speed and accuracy, especially for small defects, due to the diversity of defect types and shapes. This paper introduces an improved YOLOv5s algorithm, YOLOv5s-GSD, which enhances detection performance by integrating a Global Attention Mechanism (GAM), the SIOU loss function, and a decoupling head. The GAM enhances focus on relevant features, improving detection accuracy; the SIOU loss function replaces the GIOU loss function to optimize the vector angle of regressions, enhancing convergence speed and accuracy; and the decoupling head separates classification and regression tasks, further improving detection performance. Experimental results show the improved algorithm achieves an accuracy of 97.6% and a recognition rate of 36 frames per second, effectively reducing the miss rate of small targets.