一种改进Faster R-CNN的车辆识别算法  

An Improved Faster R-CNN Vehicle Recognition Algorithm

作　　者：朱云飞 者甜甜 刘庆华[1] ZHU Yun-fei;ZHE Tian-tian;LIU Qing-hua(School of Electronics and Information,Jiangsu University of Science and Technology,Zhenjiang 212003,China)

机构地区：[1]江苏科技大学电子信息学院,江苏镇江212003

出　　处：《软件导刊》2022年第6期25-30,共6页Software Guide

基　　金：国家自然科学基金项目(51008143);江苏省六大高峰人才项目(XYDXX-117)。

摘　　要：为了解决部分外形相似类别车辆之间的误检以及部分遮挡条件对检测的准确率影响问题,提出一种改进Faster R-CNN的车辆识别算法。该方法引入Reasoning-RCNN,结合知识图谱赋予网络推理能力,并在此基础上加入空洞卷积提高感受野,空洞空间金字塔池化强化多尺度信息提取。实验结果表明,改进的Faster R-CNN模型对于复杂场景下的车辆类型识别具有较高的敏感性,其mAP值达到93.86%,具有一定实用性。In order to solve the problem of false detection between some similar vehicles and the influence of partial occlusion conditions on the detection accuracy,an improved Faster R-CNN vehicle recognition algorithm is proposed.In this method,reasoning RCNN is introduced to give the network reasoning ability by combining with knowledge map,and on this basis,hole convolution is added to improve the receptive field,and the hole space pyramid is pooled to enhance the extraction of multi-scale information.The experimental results show that the improved Faster R-CNN model has high sensitivity for vehicle type recognition in complex scenes,and its map value reaches 93.86%,which has certain practicability.

关 键 词：深度学习 目标检测 卷积神经网络 知识图谱 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]