摘要
针对传统卷积神经网络模块体积庞大、运算量高,在体积较小、资源有限的嵌入式平台上运行效果不好,以及现有轻量化模块无法满足测量速度和测试精确度要求的问题,为此选择目前的主流目标识别算法Yolo v4进行模型轻量化,在Yolo v4模型中引入Mobilenet网络和深度可分离模块进行研究。研究结果表明,改进后不同Mobilenet网络的Yolo v4模型检测一张图片的用时均比原始Yolo v4模型减少19 ms以上,准确率都高于92%。其中以Mobilenet v3为主干特征提取网络的改进Yolo v4模型的准确率为95.12%,与原始Yolo v4模型准确率相比提高2.99%,但该模型的参数量约为Yolo v4模型的1/6,模型处理一张巡检图片用时比原Yolo v4模型减少20 ms。绝缘子作为输电线路的重要组成部分,在众多图像中更快地识别出绝缘子能为之后分析输电线路的运行情况提供帮助。
Convolutional neural network model has the disadvantages of large volume,high computation and poor performance in small and resource limited embedded platform.The existing lightweight model can not take into account the detection speed and accuracy.The mainstream target detection algorithm Yolo v4 is selected to lighten the model,and the mobilenet network and depthwise deparable convolution are used in Yolo v4 model.The results show that compared with the original Yolo v4 model,the improved Yolo v4 model of different mobilenet networks can process an image about 19 ms faster on average,and the accuracy rate can reach more than 92%.The accuracy rate of the improved Yolo v4 model with mobilenet v3 as the backbone feature extraction network is 95.13%,which is 2.99%higher than of the original Yolo v4 model.The parameter of this model is about 1/6 of Yolo v4 model,and the model can process a patrol image 20 ms faster than the original Yolo v4 model.Insulator is an important part of transmission line,The identification of insulators in many images can help to analyze the operation of transmission lines.
作者
许爱华
陈佳韵
张明文
刘浏
XU Aihua;CHEN Jiayun;ZHANG Mingwen;LIU Liu(School of Electrical and Information Engineering,Northeast Petroleum University,Daqing 163318,China)
出处
《吉林大学学报(信息科学版)》
CAS
2023年第3期545-551,共7页
Journal of Jilin University(Information Science Edition)
基金
国家自然科学基金资助项目(51774088)
黑龙江省自然基金资助项目(LH2019E016)。