SDM-YOLO: an improved YOLO model with multiscale attention for steel surface defect detection
TL;DR
The proposed SDM-YOLO framework enhances feature representation by replacing the original C2PSA module with C2PSA_SEAM in the backbone, introduces DySample-based dynamic upsampling in the neck for content-aware multi-scale feature alignment, and incorporates a multi-scale convolutional attention mechanism before the detection head to improve sensitivity to subtle, low-contrast, and morphologically varied defects.
Abstract
Surface defects generated during steel manufacturing significantly affect product quality, structural reliability, and operational safety, creating a strong demand for accurate and real-time inspection systems in industrial environments. However, existing detection approaches often struggle with subtle defect textures, complex surface backgrounds, weak visual contrast, and the trade-off between detection accuracy and computational efficiency. To address these challenges, this paper proposes SDM-YOLO, a lightweight framework for steel surface defect detection based on YOLO11n. Rather than introducing an entirely new detection architecture, the main contribution of this work lies in the coordinated integration and task-specific adaptation of complementary modules within a unified lightweight detection framework. Specifically, the proposed method enhances feature representation by replacing the original C2PSA module with C2PSA_SEAM in the backbone, introduces DySample-based dynamic upsampling in the neck for content-aware multi-scale feature alignment, and incorporates a multi-scale convolutional attention mechanism before the detection head to improve sensitivity to subtle, low-contrast, and morphologically varied defects. In addition, a normalized Wasserstein distance loss is employed to improve localization stability for small and overlapping defects without increasing inference-time parameters or computational cost. Extensive experiments on the NEU-DET and GC10-DET datasets demonstrate that SDM-YOLO achieves mAP50 scores of 81.0% and 72.3%, respectively, while attaining mAP50:95 values of 46.7% and 38.0%. The proposed framework maintains real-time performance with only 2.68 M parameters, 6.6 GFLOPs, and an inference speed of 94.5 FPS. These results demonstrate that SDM-YOLO achieves an effective balance between detection accuracy, localization precision, and computational efficiency, making it suitable for practical steel surface defect inspection applications.