摘要
PCB板表面缺陷检测是其生产中重要的环节,针对当前目标检测模型在小型工业计算平台参数多,内存负荷大等问题,提出了一种轻量化YOLOv5的检测模型。首先在主干网络采用ShuffleNetV2结构取代Conv与C3结构;其次在FPN模块加入CEM和AM模块,解决特征提取中分辨率与感受野不合的问题。最后使用PReLU激活函数代替ReLU,防止神经元崩坏。实验结果表明,改进的算法参数减少91%,每s浮点运算次数减少70%,检测精度达到95%,能够在小型计算平台完成快速精确的PCB板表面缺陷检测。
The PCB board surface defect detection is an important part of its production.Aiming at the current target detection model in small industrial computing platforms with many parameters and large memory load,a lightweight YOLOv5 detection model was proposed.First,the ShuffleNetV2 structure was used in the backbone network to replace the Conv and C3 structures.Secondly,the CEM and AM modules were added to the FPN module to solve the problem of the resolution and the receptive field in feature extraction.Finally,the PReLU activation function was used instead of ReLU to prevent the collapse of neurons.Experimental results show that the improved algorithm parameters are reduced by 91%,the number of floating-point operations per second is reduced by 70%,and the detection accuracy reaches 95%.It can complete fast and accurate PCB surface defect detection on a small computing platform.
作者
王淑青
鲁濠
鲁东林
刘逸凡
要若天
WANG Shu-qing;LU Hao;LU Dong-lin;LIU Yi-fan;YAO Ruo-tian(School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068,China;Wuhan Optoelectronics National Research Center,Huazhong University of Science and Technology,Wuhan 430074,China;School of Electrical and Automation,Wuhan University,Wuhan 430072,China)
出处
《仪表技术与传感器》
CSCD
北大核心
2022年第5期98-104,共7页
Instrument Technique and Sensor
基金
国家自然科学基金(61873195)。
