A lightweight metal surface defect detection network with hierarchical multi-branch feature extraction and group decision attention

Abstract

Metal materials are widely used in aerospace, bridge construction, and other critical applications. Surface defects such as cracks and scratches can severely undermine structural integrity and material performance, making defect detection on metal surfaces an essential industrial task. Unlike general object detection, metal surface defect detection faces unique challenges including significant scale diversity, pronounced feature ambiguity, severe data imbalance, and stringent computational resource constraints in industrial environments. To address these challenges, this study introduces LHMB-Net, a novel detection algorithm built around four key components including a hierarchical multi-branch feature extraction (HMBFE) module, a partial group decision attention (PGDA) mechanism, an HMB-head detection head, and a BoundaryIoU loss function. The lightweight hierarchical multi-scale architecture of the HMBFE module captures defect features across scales, mitigating scale diversity and feature ambiguity. The PGDA mechanism applies adaptive weighting and group decision strategies to emphasize critical features and substantially alleviate the impact of data imbalance. The HMB-Head replaces conventional convolutional structures with the HMB module to reduce model complexity while enhancing feature representation. Finally, the BoundaryIoU loss introduces boundary point distance constraints for precise localization across scales. Experimental results demonstrate that LHMB-Net outperforms current state-of-the-art methods in both detection accuracy and computational efficiency, highlighting its strong potential for practical industrial deployment.