深度学习的红外热成像电路板元器件识别研究  

Research on Infrared Thermal Imaging Circuit Board Component Recognition Based on Deep Learning

作　　者：张林鍹[1,2] 郑兴 陈飞 李名洪 邱朝洁 常乾坤 ZHANG Linxuan;ZHENG Xing;CHEN Fei;LI Minghong;QIU Chaojie;CHANG Qiankun(National Computer Integrated Manufacturing System(CIMS)Engineering Research Center,Tsinghua University,Beijing 100084,China;School of Electrical Engineering,Xinjiang University,Urumqi 830017,China;Special Equipment Inspection and Research Institute,Xinjiang Uygur Autonomous Region,Urumqi 830000,China)

机构地区：[1]清华大学国家计算机集成制造系统工程技术研究中心,北京100084 [2]新疆大学电气工程学院,乌鲁木齐830017 [3]新疆维吾尔自治区特种设备检验研究院,乌鲁木齐830000

出　　处：《电子科技大学学报》2024年第4期560-567,共8页Journal of University of Electronic Science and Technology of China

基　　金：新疆维吾尔自治区自然科学基金(2022D01C431)。

摘　　要：针对现有YOLO目标检测算法在自建数据集漏检率高、图像受复杂环境影响造成检测准确率低等问题,提出一种基于YOLO v5改进的红外图像识别算法。根据红外数据图片的独特性质,重新设计主干网络部分,引入全维动态卷积(OMNI-Dimensional Dynamic Convolution, ODConv)模块和改进坐标注意力(Coordinate Attention, CA)机制,提高模型对小目标的检测精确度并减少参数量;其次,引入解耦头(Decoupled Head, DH)模块,提高模型训练的收敛速度;最后,加入GSConv(Graph-Shifted Convolution) Slim模块,以降低模型的复杂度,提高预测速度。实验结果表明:改进后的算法模型漏检率降低40.22%,每秒浮点运算次数(Floating-point Operations Per Second, FLOPs)提升了25%,平均准确率提升了28.32%。In response to the problems of the high missing rate of existing YOLO target detection algorithms in self-built data sets and low detection accuracy caused by complex environments,proposes an improved infrared image recognition algorithm based on YOLO v5.According to the unique properties of infrared data images,the backbone network is redesigned,the OMNI-Dimensional Dynamic Convolution Module(ODConv)is introduced,and the Coordinate Attention(CA)mechanism is improved.In the meanwhile,improve the detection accuracy of small targets and reduce the number of parameters.Secondly,the Decoupled Head(DH)module is introduced to improve the convergence speed of model training.Finally,add the Graph-Shifted Convolution(GSConv)Slim module to reduce the complexity of the model and increase the speed of prediction.The experimental results show that the missed rate of the improved algorithm is reduced by 40.22%,the Floating-Point Operations per Second(FLOPs)is increased by 25%and the average accuracy is increased by 28.32%.

关 键 词：红外热成像 YOLO v5 全维动态卷积 改进解耦头 改进坐标注意力机制 GSConv Slim模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]