改进SSD的交通标志目标检测算法  

Traffic Sign Target Detection Algorithm Based on Improved SSD

作　　者：肖丹东 陈劲杰[1] XIAO Dan-dong;CHEN Jin-jie(School of Mechanical Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学机械工程学院,上海200093

出　　处：《软件导刊》2020年第5期48-51,共4页Software Guide

摘　　要：以Faster R-CNN为代表的two-stage目标检测算法检测速度慢,而one-stage目标检测算法中的SSD算法虽然检测速度快,但对交通标志类小目标的检测效果不佳。因此在SSD算法VGG16骨干网络上引入感受野块(RFB)结构,既提升检测速度又可在小目标检测上达到良好的检测精度。与此同时,为提高网络分类精度,在损失函数中加入中心损失。将SSD算法与改进的SSD算法在VOC数据集上进行训练,对比其性能可知,改进后算法mPA值达到80.7%,相比SSD300(VGG16)算法提高了3.5%。该算法在LISA traffic sign数据集上训练,在迁移学习的基础上得到的mPA值为78.4%,检测单张图像平均耗时为20.5ms,可满足实时性要求。The existing two-stage target detection algorithm represented by Faster R-CNN is slow in detection speed,while the SSD al⁃gorithm in one-stage target detection algorithm detects fast,but detects on small targets,such as traffic signs,not effectively.Therefore,the introduction of the RFB structure on the VGG16 backbone network in the SSD algorithm can achieve good detection accuracy on small target detection while taking into account the detection speed.At the same time,in order to improve the classification accuracy of the network,the center loss content is added to the loss function.SSD algorithm and the improved SSD algorithm are trained on the VOC data sets.The performance of the two algorithms are compared.The mAP value of the improved algorithm reaches 80.7%,which is 3.5%higher than that of SSD300(VGG16)algorithm.Then,based on the migration learning,the algorithm was trained on the LISA traffic sign data set,and the obtained mPA value is 78.4%,and the average time taken to detect a single image is 20.5 ms,which satisfy the requirements of real-time performance.

关 键 词：交通标志 小目标检测 RFB结构 中心损失 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]