轻量级锻件表面裂纹检测算法  

Lightweight forged part surface crack detection algorithm

作　　者：张上 许欢[1,2,3] 张岳 Zhang Shang;Xu Huan;Zhang Yue(Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering,China Three Gorges University,Yichang 443002,China;Hubei Province Engineering Technology Research Center for Construction Quality Testing Equipment,China Three Gorges University,Yichang 443002,China;College of Computer and Information Technology,China Three Gorges University,Yichang 443002,China)

机构地区：[1]水电工程智能视觉监测湖北省重点实验室,宜昌443002 [2]三峡大学湖北省建筑质量检测装备工程技术研究中心,宜昌443002 [3]三峡大学计算机与信息学院,宜昌443002

出　　处：《电子测量技术》2024年第11期123-130,共8页Electronic Measurement Technology

基　　金：省级大学生创新创业计划(S202311075047);国家级大学生创新创业训练计划(202011075013、202111075012)项目资助。

摘　　要：针对复杂场景下缺陷检测算法占用内存大、计算复杂度高和检测速度难以满足实时需求等问题,本文提出一种基于YOLOv8的轻量级锻件缺陷检测算法。首先,采集重卡转向节生产流水线探伤车间的磁粉检测图像,构建锻件表面裂纹数据集;然后,提出轻量化卷积模块GSConvns,以增强特征交互并降低计算量;同时,引入Shape-IOU损失函数,优化训练效果;最后,利用LAMP剪枝策略去除不重要的权重参数,减少模型体积并提高检测速度。实验结果表明,模型的mAP值为83.8%,参数量和计算量分别减少85.05%和80.25%,检测速度从38.7 FPS提升至65.6 FPS,显著优于其他主流算法,更适用于实时检测。在公开数据集上的测试进一步验证了其泛化能力,与基准算法相比,未剪枝的改进算法mAP值提升了2.0%。综上,本文算法能在不显著降低检测精度的前提下,大幅度提升了检测速度和资源利用效率。To address issues of high memory usage,computational complexity,and inadequate detection speed in defect detection algorithms for complex scenarios,this paper proposes a lightweight forged defect detection algorithm based on YOLOv8.First,magnetic particle inspection images from the production line of heavy truck steering knuckles were collected to construct a forged surface crack dataset.Then,a lightweight convolution module,GSConvns,was introduced to enhance feature interaction and reduce computational load.The Shape-IOU loss function was employed to optimize training performance.Finally,the LAMP pruning strategy was used to remove unnecessary weight parameters,reducing model size and increasing detection speed.Experimental results show that the model achieves a mAP of 83.8%,with parameter and computational reductions of 85.05%and 80.25%,respectively.Detection speed improved from 38.7 FPS to 65.6 FPS,significantly outperforming other mainstream algorithms,making it more suitable for real-time detection.The algorithm′s generalization capability was further verified on a public dataset,with the unpruned improved algorithm′s mAP value increasing by 2.0%compared to the baseline.In summary,this algorithm significantly enhances detection speed and resource efficiency without substantially compromising detection accuracy.

关 键 词：表面缺陷检测 YOLOv8算法 轻量化模型 损失函数 模型剪枝 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN98[自动化与计算机技术—计算机科学与技术]