Multi-scale contextual modeling and fine-grained adaptive fusion for real-time surface defect detection

Abstract Surface defect detection in steel strips requires both accurate recognition and efficient inference under complex industrial backgrounds. Existing methods still encounter difficulties in modeling scale-varied defects, preserving fine-grained defect cues, and aligning multi-level features efficiently. To address these issues, we propose CDF-YOLO, a real-time detection framework that integrates multi-scale contextual modeling and fine-grained adaptive fusion on the YOLOv12 baseline. Specifically, a Dilated Context Pyramid Bottleneck (C2f_DCPB) is introduced to enhance contextual representation through parallel dilated branches; a Fine-grained Adaptive Fusion module (FAN_Block) is embedded into the high-resolution path to strengthen local texture and structural-detail representation; and DySample is adopted to improve content-adaptive upsampling for multi-scale feature fusion. Experiments on the NEU-DET dataset show that CDF-YOLO achieves an mAP50 of 95.4% and 101.3 FPS under the adopted test protocol, while cross-dataset evaluation on DeepPCB further indicates improved generalization performance. These results suggest that CDF-YOLO provides a favorable accuracy-efficiency trade-off on public industrial defect datasets. Nevertheless, validation on larger-scale production-line data and further optimization for edge deployment remain necessary for practical industrial application.