Improved YOLOv11-Based Algorithm for Wood Surface Defect Detection

Abstract

Wood surface defect detection is a critical process in industrial quality control. However, conventional deep learning models often struggle with complex periodic wood grain interference, extreme scale variations between micro-defects and large defects, and irregular defect geometries. To address these challenges, this paper proposes WSD-YOLO, an improved object detection algorithm based on YOLOv11n. First, an Enhanced Backbone (EB) is constructed by deepening the C3k2 modules to enrich multi-level feature representations, which improves the network's capacity to extract complex defect patterns. Second, a Transformer Attention Block (TAB) with an 8-head multi-head self-attention mechanism is integrated to model long-range spatial dependencies, effectively suppressing background texture interference from wood grains. Third, a BiFPN with P2 Enhancement (BPE) extends the feature fusion network to a four-scale bidirectional architecture, utilizing a high-resolution P2 branch (stride 4) to capture fine spatial details of micro-defects such as cracks and resin spots. Finally, a Decoupled Multi-Scale Head (DMH) separates the classification and localization branches to alleviate optimization conflicts caused by irregular defect boundaries. Experimental results on a custom dataset of 4,000 images across 7 defect categories demonstrate that WSD-YOLO achieves an [email protected] of 74.34% and an [email protected]:0.95 of 39.24%, outperforming the YOLOv11n baseline by 0.82% and 1.51%, respectively. Notably, the detection precision reaches 72.18%, a 3.39% improvement over the baseline, confirming the model's superior capability in reducing false positives in complex industrial environments.