面向焊点缺陷的轻量化YOLOv9检测算法  

Lightweight YOLOv9 Detection Algorithm for Solder Joint Defects

作　　者：刘兆龙 曹伟 高军伟[1,3] LIU Zhaolong;CAO Wei;GAO Junwei(School of Automation,Qingdao University;Qingdao International Airport Group Co.LTD;Shandong Key Laboratory of Industrial Control Technology)

机构地区：[1]青岛大学自动化学院 [2]青岛国际机场集团有限公司 [3]山东省工业控制技术重点实验室

出　　处：《仪表技术与传感器》2025年第2期116-121,共6页Instrument Technique and Sensor

基　　金：山东省自然科学基金(ZR2019MF063)。

摘　　要：针对当前PCB焊点缺陷检测中存在计算量大的问题,提出一种基于改进YOLOv9的轻量化目标检测算法Sim-YOLOv9-c。首先,通过去掉采样倍数较大的检测头,重新设计PGI辅助可逆分支与主干网络进行连接,减少模型复杂程度。引入幻影卷积(GhostConv)代替常规卷积,减少计算量。通过在GELAN网络中集成无参注意力机制(SimAM),生成可靠梯度信息。实验结果证明,改进后的Sim-YOLOv9-c模型mAP达到了93%,较原模型提高了2.1%,参数量降低了48.8%,浮点运算次数降低了22.5%,并在公开数据集验证了其有效性。Aiming at the problem of heavy computation in PCB solder joint defect detection,a lightweight target detection al-gorithm Sim-YOLOv9-c based on improved YOLOv9 was proposed.Firstly,by removing the detection head with large sampling multiple,the PGI auxiliary reversible branch was redesigned to connect with the backbone network to reduce the complexity of the model.Ghost convolution(GhostConv)was introduced to replace conventional convolution,reducing computation.By integrating an parameter-free attention(SimAM)into the GELAN network,reliable gradient information was generated.The experimental re-sults show that the mAP of the improved Sim-YOLOv9-c model reaches 93%,which is 2.1%higher than that of the original mod-el,the number of parameters is reduced by 48.8%,and the number of floating point operations is reduced by 22.5%.The effec-tiveness of the improved Sim-YOLOv9-c model was verified on the public data set.

关 键 词：焊点缺陷检测 YOLOv9 轻量化 幻影卷积 无参注意力 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]