基于深度学习的车载屏幕文本检测与识别研究  

On-board screen text detection and recognition based on deep learning

作　　者：杨伟东[1] 田永祥 万峰 王炜[2] YANG Wei-dong;TIAN Yong-xiang;WAN Feng;WANG Wei(School of Mechanical Engineering,Hebei University of Technology,Tianjin,300401,China;Automotive Engineering Research Institute of China Automotive Technology Research Center Co.LTD,Tianjin 300300,China)

机构地区：[1]河北工业大学机械工程学院,天津300401 [2]中国汽车技术研究中心有限公司汽车工程研究院,天津300300

出　　处：《光电子．激光》2021年第4期395-402,共8页Journal of Optoelectronics·Laser

基　　金：河北省教育厅重点研发项目(ZD2017215)资助项目。

摘　　要：车载屏幕文本图片显示内容丰富、背景多样性、文本信息较多且大小、方向不定等问题,导致文本检测与识别过程中出现准确率低、检测速度慢的问题,因此提出了一种基于深度学习的车载屏幕文本检测与识别的系统。该系统的文本检测模型采用改进的EAST网络,应用DenseNet网络作为特征提取网络以增强特征重用,为了改善长文本的预测效果,采用区域边界元素预测顶点位置的文本线构造方法,提出了改进的文本检测模型DenseEAST网络;针对识别框架提出了CRNN-X文本识别模型,基于CRNN模型引入深度可分离卷积降低网络参数量,同时网络加入dropout以优化训练过程。实验结果表明,本文的改进方法在公开的数据集和实际车载屏幕场景的文本检测与文本识别上均有较高的准确率,模型在实际的车载屏幕场景中的文本检测准确率为97.3%、文本识别准确率97.6%,能够满足实际场景中的使用需求。On-board screen text images are rich in content,diverse in backgrounds,large in text information,and indeterminate in size and direction,resulting in low accuracy and slow detection speed in the process of text detection and recognition.Therefore,a system for detecting and recognizing text on the vehicle′s screen based deep learning-based approach in this paper is proposed.The text detection model of the system applies an improved EAST network,and uses the DenseNet network as a feature extraction network to enhance feature reuse.In order to improve the prediction of long texts,a text line construction method for predicting vertex positions of regional boundary elements is used,and an improved text detection model of DenseEAST network is put forward.The CRNN-X text recognition model of the recognition framework is proposed,the deep separable convolution based on the CRNN model is introduced to reduce the amount of network parameters,and the network adds dropout to optimize the training process.The experimental results show that the improved method for text detection and text recognition in the open data set and actual on-board screen scenes d has high accuracy.The accuracy of text detection and text recognition of the model in the actual vehicle screen scene is 97.3%and 97.6%respectively,which can meet the use requirements in the actual scene.

关 键 词：文本检测 文本识别 卷积神经网络 计算机视觉 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]