DINO-prior guided detector for limited-label cross-process steel surface defect detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Cross-process steel surface defect detection remains challenging because defect morphology, background texture, imaging conditions, and process-specific surface patterns vary substantially across manufacturing processes, whereas target-domain bounding-box annotations are often limited. Conventional detector-centric improvements may therefore struggle to distinguish true defects from process-dependent normal textures. To address this problem, we propose DINO-Prior Guided Detector (DPG-Det), which uses frozen DINOv2 representations to model cross-process surface normality. The rationale is that self-supervised DINOv2 patch descriptors provide a generic metric space that preserves local structural similarity, while prototypes constructed from steel-surface patches calibrate this space to the task-specific normality distribution. DPG-Det mines pseudo-normal patches from box-excluded regions, constructs a normality prototype bank, and converts patch-to-prototype deviations into spatial anomaly priors. These priors are injected into multi-scale detector features through anomaly-prior feature modulation, providing category-agnostic abnormality guidance. On NEU-DET and GC10-DET, DPG-Det improves mAP@0.5 over D-FINE-X and Dome-DETR-L by 1.9 and 6.0 percentage points, respectively. Under the NEU-DET $$\rightarrow $$ GC10-DET and GC10-DET $$\rightarrow $$ NEU-DET settings with 10% target labels, it further outperforms D-FINE-X by 8.5 and 7.7 percentage points, respectively. The complete framework contains 89.70M parameters and requires 311.8G FLOPs, with a measured latency of 12.70 ms per image under batch-size-one FP16 inference. Ablation studies and visualizations confirm that the proposed normality prior provides transferable spatial guidance, although this robustness is achieved at the cost of additional foundation-encoder computation.