摘要
本文概括性的介绍了图文识别所涉及的技术。首先介绍了图文识别的背景知识,包括应用领域、技术难点及挑战和系统实施流程等;其次介绍了图文识别技术的预处理方法及流程,包括旋转校正、线检测、特征匹配、字符轮廓提取及分割、OCR识别流程;接着介绍了图文识别过程中常用的特征提取基础网络和检测网络,以及它们的场景适配问题;然后介绍了近年来出现的各种图文检测深度学习网络、图文识别深度学习网络、端到端图文检测与识别深度学习网络,并分析了各类检测和识别网络的网络架构、算法思路及其特点;最后介绍了公开的图文识别训练、测试数据集以及不同算法的性能比较。
This paper gives a general introduction for the technology of image to text recognition.Firstly,the background of the image to text recognition is introduced,including application scenarios、technical difficulties and challenges、and system implementation process.Secondly,the preprocessing methods and processes of image to text recognition technology are introduced,including rotation correction、line detection、feature matching、extraction and segmentation of the character contour、and the whole processing of the OCR(Optical Character Recognition).Thirdly,we introduce the basic feature extraction network and the detection network framework commonly used in the process of image to text recognition;also,we discuss about the problem of scene adaptation when they are applied to the task of image to text recognition.Then,we introduce the various text detection deep learning network,text recognition deep learning network,end-to-end text detection and recognition network that have emerged in recent years;at the same time,we analyze the algorithm ideas and characteristics of various detection and recognition networks.Finally,we list the open data sets used in the domain of image to text recognition and performance comparison of different algorithm.
作者
牛小明
毕可骏
唐军
NIU Xiaoming;BI Kejun;TANG Jun(Sichuan Changhong Electric Co.,Ltd.,Software&Service Center,Mianyang 621000,China)
出处
《中国体视学与图像分析》
2019年第3期241-256,共16页
Chinese Journal of Stereology and Image Analysis
关键词
图文检测
文本识别
端到端识别
image to text detection
text recognition
end to end recognition
作者简介
牛小明(1983-),男(汉),硕士,四川长虹电器股份有限公司资深专家。E-mail:xiaoming1.niu@changhong.com
