Multi-Scale Deformable Transformer with Iterative Query Refinement for Hot-Rolled Steel Surface Defect Detection

Accurate and efficient detection of small and complex surface defects on hot-rolled steel plates remains a significant challenge in industrial quality assurance. Current deep learning detectors often exhibit limitations in detection accuracy and training convergence speed, particularly for small objects, which limits their practical deployment in real-time industrial inspection systems. To overcome these deficiencies, this paper proposes a multi-scale deformable transformer iterative query refinement network (MDT-Net). MDT integrates three key innovations: a Swin Transformer backbone for robust multi-scale feature representation, a deformable attention mechanism to significantly reduce computational complexity and accelerate convergence, and an iterative bounding box refinement strategy for precise localization. Extensive experiments on the NEU-DET dataset demonstrate MDT's superior performance, achieving 82.7% <i>mAP</i>₅₀. Crucially, MDT significantly outperforms other mainstream detectors in small object detection, reaching an <i>mAP</i>_50:95 of 0.55, and exhibits remarkably faster training convergence. These findings confirm MDT's effectiveness and robustness for accurate and efficient steel surface defect detection, thereby providing a crucial tool for enhancing sensor-based quality control and offering a promising solution for industrial quality management.