面向多复杂场景环境的敞车车号辨识研究
Research on coding identification of a convertible car in a complex environment
薛峰 1于国丞 1李世杰 1凌烈鹏 1张峰峰 2陈峰炜2
作者信息
- 1. 中国铁道科学研究院集团有限公司 铁道建筑研究所,北京 100081
- 2. 苏州大学 机电工程学院,江苏 苏州 215021
- 折叠
摘要
针对现有敞车车号定位识别方法存在的环境适应性差、定位和识别精度低的问题,本文提出一种面向多种复杂环境下的敞车车号精准定位和识别的方法.搭建融合多尺度特征信息的敞车车号定位模型框架,在此基础上,融合多尺度金字塔特征进行深度可分离卷积的敞车车号特征提取网络设计.提出基于改进卷积循环神经网络的车号定位识别模型,主要针对识别网络模型结构进行设计.通过不同环境下采集的敞车车厢图片对本文提出的方法进行验证.结果表明:本文提出的车号定位方法的准确率为 0.94,车号识别的准确率为 0.97.
Abstract
Existing methods for locating and recognizing the number of convertible cars have problems with poor en-vironmental adaptability and low accuracy of location and recognition.In this study,a method for the accurate posi-tioning and recognition of a convertible car number in complex environments was presented.A framework of the model for the location of the convertible car number fused with multiscale feature information was built.On this ba-sis,the features of the multiscale pyramid were fused to design the car-number feature extraction network of a con-vertible car with a deep separable convolution.In addition,a vehicle number location recognition model based on the improved convolutional recurrent neural network(CRNN)was proposed,which was mainly designed for the structure of the recognition network model.The proposed method was verified using images of the convertible car compartment collected in different environments.The results reveal that the accuracy of the proposed vehicle num-ber location method is 0.94,and that of vehicle number recognition is 0.97.
关键词
车号定位/深度可分离卷积/特征提取/改进卷积循环神经网络/特征金字塔/字符识别/铁路货运/深度学习Key words
vehicle number positioning/depthwise separable convolution/feature extraction/improved convolution-al recurrent neural network(CRNN):characteristic pyramid/character recognition/railway freight/deep learning引用本文复制引用
基金项目
中国铁道科学研究院集团有限公司科研开发基金(2022YJ099)
中铁科学技术开发有限公司基金(2022ZT05)
出版年
2024
国家科技期刊平台
NSTL
万方数据