面向木材缺陷的全局与局部协同感知检测方法

Global and Local Synergistic Perception Method for Wood Defect Detection

摘要: 针对木材表面缺陷检测中存在的缺陷形态多变、微小缺陷易漏检以及复杂背景纹理干扰等核心挑战。文章提出了一种全局与局部协同感知的检测方法。该方法以Swin Transformer作为主干网络，利用其移位窗口注意力机制构建层次化特征金字塔，有效捕获图像的全局上下文信息与长距离依赖关系；与之协同，构建CTGFusion颈部网络，该网络中，MSF模块通过轻量化MetaFormer架构与深度可分离卷积实现多尺度特征的高效融合，CWA模块则利用具备空间感知能力的门控机制对细微缺陷特征进行自适应增强，有效抑制了特征丢失与平滑问题。实验结果表明，本方法在精确率、召回率和mAP50上分别达到0.895、0.874和0.916，性能优于多个主流检测模型。在微小、低对比度缺陷的检出与复杂背景适应方面表现优异，为木材表面缺陷的自动化质检提供了可靠解决方案。

Abstract: Aiming at the core challenges in wood surface defect detection, such as significant morphological diversity, the vulnerability of micro-defects to missed detection, and interference from complex background textures, this paper proposes a detection method based on global and local synergistic perception. The method adopts Swin Transformer as the backbone network, utilizing its shifted window attention mechanism to construct a hierarchical feature pyramid, which effectively captures global contextual information and long-range dependencies of images. In synergy with the backbone, a CTGFusion neck network is developed. Within this network, the Multi-Scale Fusion module achieves efficient integration of multi-scale features through a lightweight MetaFormer architecture and depthwise separable convolutions. Meanwhile, the Channel-Wise Attention module utilizes a gating mechanism with spatial awareness to adaptively enhance fine defect features, effectively suppressing feature loss and smoothing issues. Experimental results demonstrate that the proposed method achieves a precision of 0.895, a recall of 0.874, and an mAP50 of 0.916, outperforming several mainstream detection models. The proposed method exhibits superior performance in detecting tiny, low-contrast defects and adapting to complex backgrounds, providing a reliable solution for the automated quality inspection of wood surface defects.