基于中层特征的细粒度的车型识别  

Fine-grained vehicle type recognition based on mid-level features

作　　者：宋岩贝 魏维[1] 何冰倩 SONG Yan-bei;WEI Wei;HE Bing-qian(School of Computer Science,Chengdu University of Information Technology,Chengdu 610255,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610255

出　　处：《计算机工程与设计》2020年第6期1708-1713,共6页Computer Engineering and Design

基　　金：四川省教育厅重点科研基金项目(17ZA0064)。

摘　　要：为提高细粒度车型识别的准确率,提升智能停车场、智能交通监管系统的可靠性,针对低层特征在车型识别中精确不高的问题,提出一种基于中层特征的细粒度分类算法。其核心是使用筛选算法筛选中层特征,使得筛选后特征具有较高的表示性,提高识别的准确率。使用Adaboost算法进行车脸定位,减少后期的计算量,去除干扰因素。该算法无需GPU等计算资源,方便部署。与BOW、SPM、CNN等通用的分类模型相比,其准确率有较大提升。在大众数据集中的实验结果表明,其平均准确率为95.65%,平均耗时为0.82 s。To improve the accuracy of fine-grained vehicle recognition and the reliability of intelligent parking lot and intelligent traffic supervision system,a fine-grained classification algorithm based on middle-level features was proposed to solve the problem that low-level features are not accurate in vehicle recognition.The core of this algorithm was to filter middle-level features using screening algorithm,so that the screened features showed high representativeness to improve the accuracy of recognition.The Adaboost algorithm was used to locate the car face to reduce the computational complexity and remove the interference factors.This algorithm did not need GPU and other computing resources,and was easy to deploy.Compared with the general classification models such as BOW,SPM and CNN,the accuracy is greatly improved.Experimental results on popular datasets show that the average accuracy of the algorithm is 95.65%and the average time consumed is 0.82 s.

关 键 词：车型识别 词包算法 图像分类 细粒度分类 中层特征 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]