Transformer-based wood surface defect detection with multi-scale feature fusion and density-aware enhancement

Wood surface defect detection is critical for automated quality control in the wood processing industry, yet it remains challenging due to complex grain textures, large-scale variations in defect appearance, and the prevalence of low-contrast or weak-boundary defects. To address these challenges, this paper proposed PID-DEIM, a lightweight transformer-based end-to-end detection framework. The framework integrated three key components: a partial convolution (PConv)-based backbone to reduce redundant computation in repetitive wood-grain backgrounds, an Iterative feature fusion diffusion network (IFFDN) to enhance cross-scale feature interaction, and a density-aware feature enhancement module (DAFEM) to improve sensitivity to small and inconspicuous defects. Extensive experiments on the wood surface defect dataset demonstrated that PID-DEIM achieved a mean average precision (mAP@0.5) of 86.5% and an inference speed of 85.8 FPS. Compared with the baseline DEIM, the proposed method improved mAP by 2.3 percentage points while reducing parameters, FLOPs, and memory consumption by 23.1%, 20.7%, and 22.9%, respectively. Generalization experiments on three public industrial datasets (NEU-DET, OULU-DET, and PCB) further confirmed the robustness and transferability of the proposed approach across different materials and defect types. These results indicated that PID-DEIM provided an accurate, efficient, and generalizable solution for real-time industrial surface defect inspection.