Objective Surface defect detection in wood boards is an important part of quality control in wood processing. However, the diverse morphologies of defects and their similarity to background wood-grain features lead to both missed detections of slender cracks and false detections in wood-grain regions. To address these challenges, this study proposes a structure-aware YOLO11-DWW detection model to improve the accuracy of defect recognition and localization under complex texture conditions.

Method A dataset comprising 3,930 original images covering four defect categories (live knots, dead knots, knots with cracks, and cracks) was constructed using images of wood boards sourced from poplar plantations in Qiqihar and a subset of a publicly available wood surface defect dataset. Based on YOLO11n, Directional Perception Texture Suppression Convolution (DPTSConv) was introduced to enhance the directional structural representation of slender cracks and suppress interference from regular wood-grain textures. Wood Structure Disruption-Focused Attention (WSDFAttention) was embedded to strengthen responses to regions with disrupted texture continuity. The Wood Structure-Aware Complete IoU (WSCIoU) loss was adopted to constrain bounding box regression in terms of aspect ratio consistency and effective coverage. Model effectiveness was evaluated through module ablation experiments, comparisons with mainstream object detection models, and comparisons with alternative modules of the same type. Detection results, Grad-CAM heatmaps, and intermediate feature visualizations were used to analyze responses in defect regions. Additional comparisons on a public dataset, tests under complex operating conditions, and deployment tests on different hardware platforms were conducted.

Result On the validation set, YOLO11-DWW achieved values of 0.865, 0.781, 0.837, and 0.555 for precision, recall, mAP50, and mAP50–95, respectively, exceeding those of YOLO11n by 5.6, 2.7, 5.2, and 7.1 percentage points. Compared with YOLO11n, YOLO11-DWW reduced the parameter count and floating-point operations by approximately 9.0% and 4.8%, respectively. On the Jetson Orin Nano, YOLO11-DWW achieved an inference speed of 27.12 frames/s using TensorRT FP16, approximately 1.79 times that achieved by the same model using PyTorch FP32.

Conclusion YOLO11-DWW enhanced the structural feature representation of slender defects and reduced false and missed detections against complex wood-grain backgrounds. It improved detection accuracy while maintaining low model complexity, providing a technical reference for intelligent wood board inspection and quality sorting.