Multi-Scale attention and hierarchical feature extraction-based network for metal surface defect detection

Surface defects in metal components hold significant implications for product quality and reliability, yet conventional methods suffer from the challenge of efficient and accurate detection. To address this, a novel detection network (YOLO-MAH) based on multi-scale attention and hierarchical feature extraction is proposed in this paper. First, a tiered multi-scale attention (TMA) module is constructed to enhance detail perception through the information operation method of multi-level feature progressive stacking. Second, a hierarchical feature extraction (HFE) module is developed, employing multi-scale convolutions and shortcut to establish a parallel multi-branch structure for cross-scale feature extraction while minimizing original information loss. Finally, an expanse-aware spatial pyramid pooling (EA-SPP) module is designed, leveraging multi-stage convolutional integration to expand the receptive field and facilitate multi-scale feature fusion. Experimental results on three public datasets and one self-built dataset demonstrate that YOLO-MAH achieves superior performance, with mean average precision (mAP50) improvements of 1.3%-7.7% over baseline models. Further comparisons with state-of-the-art models validate the effectiveness and generalization capability of YOLO-MAH, exhibiting its strong potential for industrial inspection applications.