DCAFNet: Deformable Context-Aware Fusion Network for Steel Surface Defect Detection

Abstract Detection of surface defects on steel is an important issue in intelligent quality control in industrial manufacturing. To address the challenges of irregular defect morphologies, complex textures, and multi-scale characteristics in steel surface defect detection, this paper proposes DCAFNet, a deformable context-aware fusion network for industrial defect inspection. In particular, we design a deformable contextual information modeling mechanism to better capture the geometric variability and contextual continuity of defects. This mechanism is responsible for the core feature extraction process. A lightweight channel-selective split attention module is introduced to enhance defect-relevant channels while retaining a complementary feature branch with limited additional computational cost. Furthermore, the detection head from YOLOv11 is adopted, incorporating depthwise separable convolutions to strike an effective balance between efficiency and localization accuracy. Experiments show that DCAFNet achieves an mAP50 of 78.7% on NEU-DET, representing a 3.3 percentage-point improvement over the baseline model. Evaluations on GC10-DET and PKU-Market-PCB further support the multi-dataset applicability and cross-scenario architectural robustness of DCAFNet.