GAM-WIoUv3: High-Compatibility Optimization for Small-Sample Defect Detection

Industrial defect detection faces severe challenges in data-scarce scenarios. Machine vision methods are commonly used for detecting defects on textile surfaces. Velcro (a common type of fiber fastener) poses high demands on detection algorithms due to its complex structure and diverse defects. Existing detection methods are often limited by insufficient training samples, resulting in poor model generalization capabilities. This paper addresses the challenge of Velcro defect detection in small-sample scenarios by proposing an improved model based on YOLOv8n. By incorporating an attention mechanism and the Wise-IoUv3 loss function, the model’s ability to process small-sample data is significantly enhanced. Experiments show that when the GAM attention module is combined with WIoUv3, detection performance is improved without loss (mAP improved by 1.2%), while other attention modules (such as CBAM) may experience performance degradation. On the self-built hook-and-loop fastener defect dataset, the proposed method achieves a mAP value of 0.826, significantly outperforming the baseline model. This study provides a plug-and-play, highly compatible optimization solution for resource-constrained industrial vision detection systems.