基于改进YOLOv8网络模型的芯片BGA缺陷检测  

Defect Detection of Chip BGA Based on Improved YOLOv8 Network Model

作　　者：陈澍元 王鸣昕 赵嘉宁 蒋忠进[1] Chen Shuyuan;Wang Mingxin;Zhao Jianing;Jiang Zhongjin(School of Information Science and Engineering,Southeast University,Nanjing 211100,China;China Electronics Corporation Pengcheng Intelligent Equipment Co.,Ltd.,Nanjing 211102,China)

机构地区：[1]东南大学信息科学与工程学院,南京211100 [2]中电鹏程智能装备有限公司,南京211102

出　　处：《半导体技术》2025年第4期378-384,392,共8页Semiconductor Technology

摘　　要：为提高芯片球栅阵列(BGA)缺陷检测的精度和效率,提出一种基于改进YOLOv8网络模型的芯片BGA缺陷检测方法。该方法以常规YOLOv8网络模型为基础,在骨干网络中引入双层路由注意力(BRA)机制,以增强模型捕捉长程依赖和复杂特征的能力;在检测头网络中引入双标签分配策略,从而省略耗时的非极大值抑制(NMS)运算,以提高模型检测速度;此外,引入Focaler-WiseIoU边框回归损失函数,以提高模型的边框回归精度。进行了大量基于实测数据的BGA缺陷检测实验,实验结果验证了改进YOLOv8模型的有效性,其总体性能优于三种对比模型。相比常规YOLOv8模型,改进YOLOv8模型的精确率提高了1.2%,召回率提高了3.9%,平均精度均值mAP_(50)提高了2.9%,mAP_(50~95)提高了2.2%,每秒处理帧数(FPS)提高了33.8%。A chip ball grid array(BGA)defect detection method was proposed based on the improved YOLOv8 network model to increase the precision and efficiency of defect detection for BGA packages.The method was based on the conventional YOLOv8 network model,and the bi-level routing attention(BRA)mechanism was introduced into the backbone network to strengthen the ability of the model to capture long-range dependencies and complex features.A dual-label assignment strategy was introduced into the detection head network,eliminating the time-consuming non-maximum suppression(NMS)operations and accelerating the detection speed of the model.Additionally,a Focaler-Wise IoU bounding box regression loss function was introduced to enhance the bounding box regression accuracy of the model.Extensive experiments of BGA defect detection were conducted based on measured data,and the effectiveness of the improved YOLOv8 model was validated by the experimental results,demonstrating superior overall performances of the improved YOLOv8 model over the three contrast models.Compared with the conventional YOLOv8 model,the improved YOLOv8 model achieves a 1.2%increase in precision,a 3.9%increase in recall,a 2.9%increase in mean average precision mAP_(50),a 2.2%increase in mAP_(50~95),and a 33.8%increase in the frames per second(FPS).

关 键 词：球栅阵列(BGA) 缺陷检测 深度学习 YOLOv8 注意力机制 

分 类 号：TN407[电子电信—微电子学与固体电子学]