Fabric Defect Detection Algorithm Design Based on Improved YOLOv5 Model

This study evaluates fabric-specific anchor adaptation and Coordinate Attention (CA) within YOLOv5s for six defect categories. The detection study uses 7921 images from Alibaba Tianchi and laboratory acquisition. Nine anchor shapes are derived using a 1-IoU dissimilarity, and SE, CBAM, and CA are compared at a fixed backbone location. Anchor adaptation increases [email protected] from 85.04% to 87.42%. Combining the adapted anchors with CA achieves 91.35% precision, 83.14% recall, and 87.05% [email protected], improving recall by 1.60 percentage points over the baseline. Its [email protected] exceeds YOLOv8n, YOLOv9t, YOLOv10n, and YOLO11n by 1.64, 5.05, 4.46, and 10.26 percentage points, respectively. Comparative anchor-clustering experiments on the same fabric dataset converge in all 30 runs for each of the two initialization schemes, with mean best possible recall above 99.8%. The proposed configuration contains 7.103 million parameters and reaches 127.55 FPS for the measured GPU processing pipeline at 640 × 640 resolution and batch size one. The results support fabric-specific configuration selection through detection ablations, repeated anchor analysis, and measured computational costs.