«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

31

期数:

2021年03期

页码:

89-94

栏目:

图形与图像

出版日期:

2021-03-10

文章信息/Info

Title:

Research on Fine Identification of Vehicle Type Based on Capsule Neural Network

文章编号:

1673-629X(2021)03-0089-06

作者:

程换新¹; 郭占广¹; 程 力²; 刘文翰¹; 张志浩¹

1. 青岛科技大学 自动化与电子工程学院,山东 青岛 266061;

2. 中国科学院新疆理化技术研究所,新疆 乌鲁木齐 830001

Author(s):

CHENG Huan-xin¹; GUO Zhan-guang¹; CHENG Li²; LIU Wen-han¹; ZHANG Zhi-hao¹

1. School of Automation and Electronic Engineering,Qingdao University of Science and Technology,Qingdao 266061,China;

2. Xinjiang Technical Institute of Physics and Chemistry,Chinese Academy of Sciences,Urumqi 830001,China

关键词:

人工智能; 胶囊神经网络; 车型精细识别; 智能交通; 深度学习; CapCar 模型

Keywords:

artificial intelligence;  capsule neural network;  fine identification of vehicle type;  smart transportation;  deep learning; CapCar model

分类号:

TP391.9

DOI:

10. 3969 / j. issn. 1673-629X. 2021. 03. 015

摘要:

车辆型号精细识别在智能交通系统、涉车刑侦案件侦破等方面具有十分重要的应用前景。针对车辆型号种类繁多、部分型号区分度小等带来的车辆型号精细分类困难的问题,提出一种基于胶囊神经网络(capsule network, CapsNet)的车型图像识别模型 CapCar。 以 CompCars 数据集作为样本,首先通过加权平均值法进行图像的灰度化处理,减少数据集训练计算量,提高模型的训练速度。 然后通过胶囊神经网络提取车型图像的全部特征和局部特征,实现车型分类识别。相较于现有的车型精细识别方法,该方法在提高识别精度的同时,有效压缩模型参数规模。在基准数据集 CompCars 下进行大量实验结果表明, CapCar 模型进行车辆精细识别精度可达 98. 89%,其识别率高于一些其他经典的网络模型。CapCar 模型参数大小仅为 6. 3 MB。 该算法具有一定的先进性。

Abstract:

The fine identification of vehicle type has an important application prospect in intelligent transportation syste-ms and the investigation of criminal cases involving vehicles. Aiming at the difficulty of fine classification of vehicle models caused by the wide variety of vehicle models and the small degree of differentiation of some models,we propose a vehicle image recognition model CapCar based on capsule network (CapsNet). Taking the CompCars data set as a sample,first of all,the image gray processing is performed by the weighted ave-rage method to reduce the training calculation amount of the data set and improve the training speed of the model. Then all the features and local features of the vehicle model image are extracted through the capsule neural network to realize fine identification of vehicle type. Compared with the existing precise identification method of vehicle models,the proposed method effectively reduces the model parameter scale while impro-ving the identification accuracy. The results of a large number of experiments under the benchmark dataset CompCars show that the CapCar model can achieve a fine vehicle recognition accuracy of 98. 89% ,and its recognition rate is higher than some other classic network models. The CapCar model parameter size is only 6. 3 MB. The proposed algorithm has a certain degree of advancement.

常用功能

更新日期/Last Update: 2020-03-10