Abstract: The quality of metal materials directly affects the safety of industrial products. However, traditional defect detection algorithms on metal material quality cannot achieve high-precision real-time detection. To address it, a metal surface defect detection framework called you only look once v7-bi concatenation adaptive (YOLOv7-BCA) has been proposed. The YOLOv7-BCA framework first uses dynamic sparse sampling for feature extraction, enhancing the network's ability to extract fine-grained features. Next, a feature enhancement module is designed to achieve feature fusion and enhancement for three adjacent layers, improving the accuracy of localization information. Finally, combining the idea of adaptive spatial feature fusion, the semantic information from deep network feature maps and the positional information from shallow network feature maps are effectively integrated. Comprehensive and detailed experimental results demonstrate that the proposed YOLOv7-BCA achieves an average detection accuracy of 75.9%, a 4.7% improvement over the original model. YOLOv7-BCA exhibits significant advantages in metal surface defect detection, it can achieve highly accurate localization of metal surface defects and demonstrating strong real-time performance.

Key words: defect detection, sparse attention mechanism, adaptive spatial feature fusion