基于改进YOLOv8算法的焊缝缺陷检测研究

Research on Weld Defect Detection Based on Improved YOLOv8 Algorithm

摘要: 缺陷识别与分类是机器视觉焊缝缺陷检测流程的核心收尾环节。针对焊缝缺陷(如气孔、夹渣、裂纹、未熔合)特征微弱、易被噪声与背景纹理淹没的识别难题，文章提出一种基于YOLOv8算法改进的YOLO-CCS模型。首先使用CSWinTransformer替换了原骨干网络的特征提取模块，之后对其进行结构化通道剪枝，降低模型参数量和计算量。其次在FPN-PAN网络中融合双注意力SimAM和CBAM形成AttentionFPNPAN颈部模块，起到抑制焊接背景干扰和强化微小缺陷(Φ0.2~2 mm)特征的作用。同时新增P2微小缺陷检测头，移除了原来的P5大目标检测头，解决了对微小缺陷出现误检漏检的问题。结果表明，训练后的精确率为92.3%，召回率为90.2%，mAP@0.5 (%)达到90.3%，效果优于其他主流算法。

Abstract: Defect detection and classification constitute the critical concluding phase in machine vision-based weld defect inspection processes. To address the challenge of weak feature recognition for weld defects (e.g., porosity, slag inclusion, cracks, and incomplete fusion) that are easily obscured by noise and background textures, this study proposes an improved YOLO-CCS model based on the YOLOv8 algorithm. First, the original backbone network’s feature extraction module is replaced with CSWinTransformer, followed by structured channel pruning to reduce model parameters and computational complexity. The FPN-PAN network integrates dual attention mechanisms (SimAM and CBAM) in the AttentionFPNPAN neck module, effectively suppressing welding background interference while enhancing subtle defect (Φ0.2~2 mm) detection. Additionally, a new P2 micro-defect detection head replaces the original P5 large-object detection head to resolve false detection and missed detection issues. Experimental results demonstrate a trained precision rate of 92.3%, recall rate of 90.2%, and mAP@0.5% score of 90.3%, outperforming other mainstream algorithms.