PSDW-YOLO: a pixel-preserving shallow-deformable network for steel surface defect detection

Abstract

Steel surface defects directly affect product quality, structural reliability, and downstream manufacturing safety. Although YOLO-based detectors provide high inference speed, steel surface defect detection remains challenging because many defects in the NEU-DET dataset are small, low-contrast, irregularly shaped, and difficult to localize accurately. To address these problems, this paper proposes PSDW-YOLO, a pixel-preserving, shallow-feature, deformable, and WIoU-optimized detection framework based on YOLOv11n. Specifically, SPD-Conv is introduced into the backbone to reduce information loss caused by conventional strided downsampling. A P2 shallow detection head is constructed to enhance high-resolution spatial representation for tiny defects. DCNv4 is integrated into selected C3k2 modules to improve adaptive modeling for irregular and slender defect morphologies, and Wise-IoU is adopted to improve bounding box regression. Experiments on the NEU-DET dataset show that PSDW-YOLO achieves 72.3% mAP@0.5 and 39.8% mAP@0.5:0.95, outperforming YOLOv11n by 2.6 and 5.5 percentage points, respectively, while maintaining 89 FPS on an NVIDIA RTX 4090. The results demonstrate that the proposed method improves detection accuracy and localization robustness while preserving real-time performance.