A lightweight wood surface defect detection method integrating bivariate features and attention mechanism
Abstract
To address the challenges of balancing detection accuracy and real-time performance in identifying diverse wood surface defects with varying sizes and complex textures, this study proposes a lightweight defect detection method integrating YOLOv8 and MobileNetV4-Conv-S. First, MobileNetV4-Conv-S serves as the backbone network, employing a universal inverted bottleneck structure to reduce model parameters and computational complexity while enabling efficient feature extraction. Second, a bidirectional feature fusion architecture enhances multi-scale feature representation through top-down and bottom-up information exchange mechanisms, improving detection performance for fine-scale defects. Additionally, a dual attention mechanism weights features across spatial and channel dimensions, enhancing the model's sensitivity to critical regions. Experiments on a dataset comprising 5,466 images with six defect categories demonstrated that the proposed model achieves high detection accuracy (mAP: 95.5%, recall: 94.2%, precision: 96.8%) while maintaining lightweight architecture. The results confirm the method's superior performance and real-time capabilities for wood surface defect detection, showcasing significant potential for engineering applications.