DSAT: a dynamic sparse attention transformer for steel surface defect detection with hierarchical feature fusion
Shouluan Wu, Hui Yang, Liefa Liao, Chao Song, Yating Fang, Jianglong Fu, Tan Li
Abstract
The rapid development of industrialization has led to a significant increase in the demand for steel, making the detection of surface defects in steel a critical challenge in industrial quality control. These defects exhibit diverse morphological characteristics and complex patterns, which pose substantial challenges to traditional detection models, particularly regarding multi-scale feature extraction and information retention across network depths. To address these limitations, we propose the Dynamic Sparse Attention Transformer (DSAT), a novel architecture that integrates two key innovations: (1) a Dynamic Sparse Attention (DSA) mechanism, which adaptively focuses on defect-salient regions while minimizing computational overhead; (2) an enhanced SPPF-GhostConv module, which combines Spatial Pyramid Pooling Fast with Ghost Convolution to achieve efficient hierarchical feature fusion. Extensive experimental evaluations on the NEU-DET and GC10-DE datasets demonstrate the superior performance of our approach. On the NEU-DET dataset, DSAT achieved an accuracy of 92.55%, with a mean Average Precision (mAP) of 83.14% at an IoU of 0.5 and 47.37% at an IoU of 0.5:0.95, significantly outperforming existing methods. Similar improvements were observed on the GC10-DE dataset, where the model attained an accuracy of 79.61% and a mAP of 67.27% at an IoU of 0.5, and a mAP of 34.09% at an IoU of 0.5:0.95. Comprehensive ablation studies validated the respective contributions of the DSA mechanism and the SPPF-GhostConv module, while visualization experiments demonstrated the model's enhanced capability in detecting fine-grained defects. Overall, the proposed DSAT model exhibits significant advantages and potential in the field of steel surface defect detection.