An Efficient DETR-Based Framework for Small Target-Aware Industrial Surface Defect Detection

Industrial surface defect detection requires accurate localization of small and weak-boundary defects under tight runtime constraints for on-line inspection. This paper presents an efficient DETR-style defect detector with three components. First, we build a hybrid feature extractor by coupling a ConvNeXt-T backbone with a lightweight Feature Pyramid Network (FPN) to strengthen multi-scale representations for small and subtle defects, thereby improving detection performance in challenging industrial environments. Second, to address the high computational cost of original DETR, we adopt multi-scale deformable attention to replace the quadratic-cost global self-attention mechanism, substantially improving efficiency. Third, to improve per-class robustness on hard defect categories with negligible overhead, we incorporate a class-reweighted focal loss (focal loss with no-object down-weighting together with effective-number reweighting) for classification. Experiments on NEU-DET show that our method achieves 85.0% mAP@0.5 and 47.5% mAP@[0.5:0.95], improving over the original DETR baseline (81.0% mAP@0.5). Under the same runtime setup (NVIDIA 3090, FP16, batch size 1, 512 $\times$ 512), latency is reduced from 71.4 to 53.3 ms with 18.8 img/s throughput. These results demonstrate that our framework achieves a superior accuracy-efficiency trade-off for near real-time on-line inspection (approximately 18.8 img/s, 53.3 ms latency) in medium-speed large-scale industrial manufacturing scenarios.