基于感兴趣区域卷积神经网络的车辆检索方法  

Vehicle retrieval method based on region of interest convolutional neural network

作　　者：杨东芳 白艳宇[2] YANG Dong-fang BAI Yan-yu(Department of Information Engineering, Huanghe Jiaotong University, Jiaozuo 454950, China Department of Information Technology, College of Information and Business, Zhongyuan University of Technology, Zhengzhou 450007, China)

机构地区：[1]黄河交通学院信息工程系,河南焦作454950 [2]中原工学院信息商务学院信息技术系,河南郑州450007

出　　处：《计算机工程与设计》2017年第8期2276-2280,F0003,共6页Computer Engineering and Design

基　　金：河南省科技攻关重点计划基金项目(122102210563;132102210215);河南省高等学校重点科研项目计划基金项目(15B520008)

摘　　要：为降低车辆检索的误检率并提高检索效率,对块卷积神经网络进行优化,提出一种车辆检索方法。增加感兴趣区域输入层、感兴趣区域池化层和目标包围盒输出层,提高传统块卷积神经网络模型的运算效率和分类性能;采用优化的模型进行车辆检索,依据边缘检测、轮廓提取获取感兴趣区域,通过优化的模型提取特征和进行特征分类,输出车辆目标及其位置。实验结果表明,该方法的车辆检索误检率和检索效率优于目前主流车辆检索方法。To reduce the false positive rate of vehicle retrieval and improve the search efficiency, a vehicle retrieval method was proposed by optimizing the region-based convolutional neural network. Input layer and pooling layer of region of interest, and output layer of target bounding box were added, for improving the efficiency and performance of traditional region-based convolutional neural network. The optimization model for vehicle retrieval was used in which region of interest was obtained through edge detection and contour extraction, features were extracted and executed through the optimization model, and vehicle targets were outputted with its position. Experimental results show that the false positive rate and retrieval efficiency of vehicle retrieval using the proposed method are better than the current mainstream methods of vehicle retrieval.

关 键 词：卷积神经网络 车辆检索 最大池化 后向传播 轮廓提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]