Steel surface defect detection based on multi-scale dynamic convolution and lightweight cross-stage fusion

To address the primary challenges of inadequate multi-scale defect perception capabilities and limited computational resources on edge devices in steel surface defect detection scenarios, this article proposes a steel surface defect detection based on multi-scale dynamic convolution and lightweight cross-stage fusion (LightSDD). First, to adaptively enhance the feature representation ability of small targets, the multi-scale dynamic convolution kernel is fused with the channel attention mechanism, and a fine-grained feature fusion method based on C3k2 multi-scale dynamic convolution is proposed. Next, to reduce redundant parameters while retaining key positioning information, an interpolation alignment strategy is adopted for multi-resolution features, along with channel compression and cross-resolution feature concatenation, based on the replacement of the original you only look once version 11 (YOLOV11) detection head. This approach enables the design of a lightweight cross-stage detection head, LSCD. Later, comparative experiments on the NEU-DET and GC10-DET datasets demonstrate that LightSDD outperforms seven state-of-the-art methods. Finally, performance tests on edge devices indicate that LightSDD exhibits strong robustness and practicality. Code is available at: https://github.com/chaoszzz-zyx/LightSDD (DOI: 10.5281/zenodo.17012588 ).