用于交通标志检测的窗口大小聚类残差SSD模型  

A Residual SSD Model Based on Window Size Clustering for Traffic Sign Detection

作　　者：宋青松[1] 王兴莉 张超 陈禹 宋焕生[1] KHATTAK Asad Jan SONG Qingsong;WANG Xingli;ZHANG Chao;CHEN Yu;SONG Huansheng;KHATTAK Asad Jan(School of Information Engineering,Chang’an University,Xi’an710064,China)

机构地区：[1]长安大学信息工程学院

出　　处：《湖南大学学报（自然科学版）》2019年第10期133-140,共8页Journal of Hunan University:Natural Sciences

基　　金：国家自然科学基金资助项目(61201406,61572083);中国博士后科学基金资助项目(2019M653516)~~

摘　　要：SSD通常被认为适合于求解小目标图像检测问题,但在特征表征和检测效率两方面还存在改进空间.提出一种聚类残差SSD模型,一方面将原始SSD模型中的VGG16基础网络替换为更深的ResNet50残差网络,以改善特征表征能力.另一方面采用K-均值聚类算法取代盲目搜索机制,确定SSD中默认窗口的大小,以改善检测效率.针对德国交通标志检测数据集,模型获得了97.1%mAP和每幅图像0.07 s的检测速度.针对中国交通标志数据集,模型获得89.7%mAP和每幅图像0.08 s的检测速度.与原始SSD模型比较,本文所提模型的检测性能得到改善.Single Shot MultiBox Detector(SSD)is generally considered to be suitable for solving small target detection in images.However,its performance on feature extraction and detection efficiency is still required to be improved.A clustering residual SSD model is proposed in this paper.On one hand,in order to improve the feature extraction quality,the basic network VGG16 which consists of the original SSD model is replaced with a deeper residual network ResNet50.On the other hand,in order to improve the detection efficiency,K-means algorithm other than the blind search mechanism used in the original SSD model is exploited to find and determine the assignments of the sizes of default windows.For German traffic sign detection dataset,it obtains 97.1%mAP in detection accuracy and 0.07 s per image in detection efficiency.For Chinese traffic sign dataset,it obtains 89.7%mAP in detection accuracy and 0.08 s per image in detection efficiency.Compared with the original SSD model,the proposed model obtains the improved detection performance.

关 键 词：交通标志检测 深度学习 单拍多盒探测器(SSD) K-均值 聚类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]