A Multi-Scale Feature-Enhanced YOLOv11 for Wood Surface Defect Detection

Wood defects, including cracks, knots, and resin pockets, compromise structural integrity and product quality. This paper proposes an improved YOLOv11 detector to address complex texture interference, high miss-detection rates, and multi-class recognition challenges. MobileNetV4 is adopted as the backbone for multi-scale feature extraction, DySample dynamic upsampling is incorporated into the neck for detail-preserving feature fusion, and BiFormer bi-level routing attention is integrated into the detection head for enhanced contextual representation. Evaluated on a seven-class wood surface defect dataset, the model achieves an mAP@50 of 0.842 and Precision of 0.857, surpassing YOLOv11n by 1.5% and 3.1% respectively, and outperforming state-of-the-art methods.