Improving YOLO Detection Performance for High-Speed Steel Manufacturing via Normalized Statistical Contrast Layers

Automated surface defect detection in high-speed steel manufacturing demands robust, real-timecapable models that can reliably identify subtle, low-contrast anomalies under industrial conditions. While YOLO-based detectors offer an attractive speed–accuracy trade-off, their performance on challenging defect types—such as crazing or rolled-in scale—remains limited by insufficient local contrast representation. To address this, we propose MaxSigLayerNormalized, a novel, lightweight, and learnable contrast enhancement module derived from MaxSigLayer, which adaptively amplifies discriminative features using a statistically normalized formulation based on center, median, and mean statistics. Integrated via a plug-and-play block—MaxSigC2f—into both the backbone and neck of YOLOv8 andYOLOv11 architectures, our method improves detection sensitivity with negligible computational overhead (<2% increase in GFLOPs). Extensive experiments on the NEU-DET benchmark demonstrate consistent gains across model scales and splits, with mAP@0.5 improving from 0.726 to 0.744 and mAP@0.5 :0.95 from 0.375 to 0.399. The approach exhibits strong generalizability, robustness under stratified cross-validation, and particular efficacy on low-contrast defect classes. By combining statistical rigor with architectural compatibility, MaxSigLayerNormalized offers a practical, deployable enhancement for industrial vision systems requiring high reliability and real-time performance.