Industrial Small Object Defect Detection Method Based on YOLO11 Integrating Directional Inductive Bias and Heterogeneous Feature Reconstruction

Industrial surface defects often occupy limited spatial regions and exhibit pronounced directional structures, posing challenges for accurate detection in automated visual inspection. Existing detection methods may struggle to capture directional geometric features, distinguish tiny defects from complex backgrounds, and preserve spatial details during cross-scale feature fusion. To address these issues, this study proposes a YOLO11-based industrial surface defect detection method that integrates directional inductive bias and heterogeneous feature reconstruction. First, direction-aware Pinwheel Convolution (PConv) is introduced at the beginning of the backbone to enhance the representation of directional structures in anisotropic defects, such as elongated cracks and scratches. Second, a Multi-Branch Heterogeneous Feature Extraction (MGHFE) module is designed to integrate fine-grained directional features with contextual semantic information through parallel heterogeneous branches. Finally, a Hierarchical Scale Feature Fusion Module (HSFM) is incorporated into the neck to facilitate cross-scale feature interaction and preserve spatial details associated with small defects. Experiments on the NEU-DET dataset show that the proposed method achieves an mAP@0.5 of 80.4% and an mAP@0.5:0.95 of 46.5%, with 3.5 million parameters and 9.1 GFLOPs. Further evaluation on the self-constructed AHD-Defect dataset demonstrates the method’s effectiveness in detecting industrial surface defects under practical inspection conditions.