The Study of Improved YOLOv13-Based Method for Detection of Industrial Surface Defects

Surface defect detection is a core of industrial product quality control, vital for ensuring product reliability and production efficiency. However, due to diverse types and significant size variations of industrial surface defects—especially minute or complex ones—accurate feature extraction and efficient detection remain major challenges, and existing You-Only-Look-Once (YOLO) methods struggle to meet high-precision demands. This paper proposes a symmetry-aware YOLOv13-based industrial surface defect detection network. First, a Multi-level Feature Enhancement Module (MFEM) is designed, combining a star-shaped architecture with the CBAM attention mechanism to enhance defect feature discriminability via multi-branch feature interaction and nonlinear expression, while compensating for detail loss from multi-layer depth-wise separable convolutions (DSConv). The symmetric dual-branch structure in MFEM improves feature balance and structural consistency. Second, the Spatial Pixel Global Attention (SPGA) module is introduced to supplement detail information during feature pyramid transmission and enhance multi-scale feature fusion efficiency, while maintaining symmetric feature distribution. Third, the HyperACE module is improved using a multi-branch hypergraph structure to enhance long-range dependency modeling and local feature representation. On the GC10-DET dataset, the improved model achieved 69.6% Precision, 66.1% Recall, and 67.0% mAP@50, demonstrating superior performance while maintaining real-time capability.