改进MobileNetv3-YOLOv3交通标志牌检测算法  

Improved MobileNetv3-YOLOv3 traffic sign detection algorithm

作　　者：刘宇宸 石刚[1] 崔青[1] 刘明辉 郑秋萍 LIU Yu-chen;SHI Gang;CUI Qing;LIU Ming-hui;ZHENG Qiu-ping(College of Information Science and Engineering,Xinjiang University,Urumqi 830046,China)

机构地区：[1]新疆大学信息科学与工程学院,新疆乌鲁木齐830046

出　　处：《东北师大学报（自然科学版）》2022年第2期53-60,共8页Journal of Northeast Normal University(Natural Science Edition)

基　　金：新疆维吾尔自治区自然科学基金资助项目(2020D01C047).

摘　　要：提出了一种改进的MobileNetv3-YOLOv3算法.引入SPPNet去除重复特征,提高候选框的选取速度,加快模型的推理速度;引入CSPNet消除计算瓶颈,节省计算量;引入Focus层防止特征图信息丢失,保证模型的准确度.在CCTSDB数据集上的实验结果表明:MobileNetv3与YOLOv3算法结合最佳,改进MobileNetv3-YOLOv3算法的平均精度高达97.7%,检测速度FPS达到89,与YOLOv3算法相比较,本文算法精确度(P)、召回率(R)、平均精度值(mAP)、每秒传输帧数(FPS)都得到了提升,分别提升6%,1.8%,1.8%和15%,与最新的算法相比较本文的P值以及mAP值均取得了最优的结果.可以看出所提出的算法在极大地减少模型参数量和计算量的同时,提高了检测精度以及检测速度.Traffic sign detection is an important part of unmanned and assisted driving.In order to better meet the detection accuracy and real-time characteristics, this paper proposes an improved MobileNetv3-YOLOv3 algorithm.SPPNET is introduced to remove repeated features, improve the selection speed of candidate boxes and speed up the reasoning speed of the model.CSPNET is introduced to eliminate the calculation bottleneck and save the calculation amount.Focus layer is introduced to prevent the loss of feature map information and ensure the accuracy of the model.Experimental results on CCTSDB data sets show that the combination of MobileNetv3 and YOLOv3 algorithm is the best.The average accuracy of the improved MobileNetv3-YOLOv3 algorithm is as high as 97.7%,and the detection speed FPS is up to 89.Compared with YOLOv3 algorithm, the algorithm in this paper obtains P value, R value and mAP value, FPS has been improved by 6%,1.8%,1.8%and 15%,respectively.Compared with the latest algorithm, the P value and mAP value in this paper have obtained the optimal results.It can be seen that the algorithm proposed in this paper not only greatly reduces the number of model parameters and calculation amount, but also improves the detection accuracy and detection speed.

关 键 词：交通标志检测 MobileNet 聚焦层 跨阶段局部网络 

分 类 号：TN911.73[电子电信—通信与信息系统]