基于改进PP-YOLOv2的IC引脚焊接缺陷检测算法研究  

Research on IC Pin soldering Defect Detection AlgorithmBased on Improved PP-YOLOv2

作　　者：李娜 王学影 胡晓峰 郭斌 罗哉 LI Na;WANG Xue-ying;HU Xiao-feng;GUO Bin;LUO Zai(College of Metrology and Measurement Engineering,China Jiliang University,Hangzhou,Zhejiang 310018,China)

机构地区：[1]中国计量大学计量测试工程学院,浙江杭州310018

出　　处：《计量学报》2023年第10期1574-1581,共8页Acta Metrologica Sinica

基　　金：国家自然科学基金(52075511);浙江省科技计划项目省级重点研发计划(2021C01136);浙江省公益性技术应用研究计划(LGG21E050019)。

摘　　要：针对IC引脚焊接缺陷因目标尺寸小、引脚密集导致检测精度低等问题,提出一种基于改进PP-YOLOv2的IC引脚焊接缺陷检测算法。通过在骨干网络后引入SE注意力机制,区分特征图中不同通道的重要性,强化目标区域的关键特征,提升网络的特征提取能力。使用k-means++聚类算法产生9个聚类中心,以降低因初始聚类中心随机选择不当对检测结果所造成的误差影响。实验结果表明:改进算法对IC引脚焊接短路、缺脚、翘脚、少锡缺陷检测的平均精度分别为97.9%, 96.1%, 96.7%, 95.8%;在阈值为0.5的情况下,平均精度均值达到了96.6%,与YOLOv3、PP-YOLOv2相比,分别提高了14.9%, 5.1%。改进算法对单幅图片的检测时间为0.151 s,满足IC质检的速度要求,为IC引脚焊接缺陷检测提供了参考价值。Aiming at the low detection accuracy of IC pin soldering defects due to small target size and dense pins,an algorithm for IC pin soldering defects detection based on improved PP-YOLOv2 is proposed.By introducing SE attention mechanism behind the backbone network,the importance of different channels in the feature map is distinguished,the key features of the target area are strengthened,and the network feature extraction ability is improved.k-means++clustering algorithm is used to generate 9 cluster centers to reduce the error impact on the detection results caused by improper random selection of initial cluster centers.The experimental results show that the average accuracy of the improved algorithm for detecting the defects of IC pin soldering short circuit,missing pin,warping pin and little tin is 97.9%,96.1%,96.7%and 95.8%respectively.Under the threshold value of 0.5,the average accuracy reaches 96.6%,which is 14.9%and 5.1%higher than YOLOv3 and PP-YOLOv2 respectively.The detection time of the improved algorithm for a single picture is 0.151 s,which meets the speed requirements of IC quality inspection and provides a reference value for IC pin soldering defect detection.

关 键 词：计量学 焊接缺陷检测 IC引脚 改进PP-YOLOv2算法 SE注意力机制 k-means++ 机器视觉 

分 类 号：TB973[一般工业技术—计量学] TB96[机械工程—测试计量技术及仪器]