Abstract
Wood surface defect detection remains challenging due to interference from natural wood textures, large variations in defect scale, and difficulty in localizing small targets, all of which contribute to elevated rates of false and missed detections. To address these challenges, we propose TFC-YOLO26, a specialized detection model built upon YOLO26 with targeted feature enhancement. The texture injection (TI) module is designed to compel the model to capture discriminative features amid complex texture interference by injecting texture statistics into feature maps; the frequency–amplitude–phase attention (FAPA) module is devised to improve frequency-domain differentiation between defects and background textures by jointly fusing frequency, amplitude, and phase information; and the coordinate attention module is introduced to enhance spatial localization accuracy for small defect targets. The experimental results show that a TI coefficient of 0.5 yields the best overall performance for the TI module, and the ablation experiment of the FAPA module validates the design rationale and effectiveness of the FAPA module. On the primary benchmark dataset, TFC-YOLO26 achieves an mAP50 of 0.8008 and an mAP50-95 of 0.5804, representing improvements of 1.85% and 1.24% over the baseline, respectively, while outperforming several mainstream object detection models. Cross-dataset evaluations on additional wood and steel defect datasets further confirm the model’s strong generalizability and practical potential for industrial deployment.