Research on surface defect detection method for industrial parts based on YOLOv8-ECA

To address the challenges of surface defect detection in industrial parts, such as small defect size, diverse morphology, and susceptibility to interference, this paper proposes an improved detection framework, YOLOv8-ECA, which integrates a lightweight attention mechanism and a multiscale feature fusion structure. Based on the YOLOv8 backbone, this method introduces an efficient channel attention (ECA) module to enhance the responsiveness of key feature regions, incorporates a bidirectional feature pyramid network (BiFPN) to achieve efficient fusion of semantic information at different scales, and designs a decoupled detection head to optimize the representation capabilities of the classification and localization branches respectively. To verify the effectiveness of the proposed method, extensive experiments were conducted on a constructed industrial surface defect dataset. The results show that YOLOv8-ECA outperforms mainstream detection models in metrics such as mAP@0.5, mAP@0.5:0.95, and F1-score while maintaining high inference speed. Particularly, this method exhibits stronger robustness in typical interference scenarios such as blurring, occlusion, and reflection. The proposed method provides a deployable and interpretable solution for high-precision surface defect detection in complex industrial environments.