基于语义分割技术的任意方向文字识别  

Text recognition in any direction based on semantic segmentation

作　　者：王涛[1] 江加和[1] WANG Tao;JIANG Jiahe(School of Automation Science and Electrical Engineering,Beihang University,Beijing 100191,China)

机构地区：[1]北京航空航天大学自动化科学与电气工程学院,北京100191

出　　处：《应用科技》2018年第3期55-60,共6页Applied Science and Technology

基　　金：国家自然科学基金项目(61273141)

摘　　要：针对现有文本检测与定位方法只能处理单一方向文本行的缺点,提出了一种基于语义分割方法的用于自然图像中文本检测的新方法。首先通过对现有检测方法以及目前语义分割方法在文本行检测中的局限性分析。然后对加入矩形卷积核的全卷积网络模型进行训练,获得文本行区域的分类图。最后,通过全连接条件随机场(conditional random field,CRF)的高精度分割能力将网络前端输出的文本行区域中的文字给区分出来。该框架用于处理任意方向、语言和字体中的文本。所提出的方法在MSRA-TD500和ICDAR2015两个文本检测数据集上获得良好的分割结果且性能优越。A new method for text detection in natural images based on semantic segmentation method was proposed to solve the shortcomings of the existing text detection and location methods. Firstly, the limitations of existing detection methods and current semantic segmentation methods in text line detection were analyzed; then the full convolution network（FCN） model added with rectangular convolution kernel was trained to obtain the classification map of the text line area; finally, by the precision segmentation of the full-connection conditional random field（CRF）, the characters in the text line area of network front-end output were distinguished. The framework can be used to process the texts in any direction, language, and font. The proposed method achieved surprising segmentation results in both MSRA-TD500 and ICDAR2015 text detection datasets and achieved advanced performance.

关 键 词：自然图像 任意文本行 文本检测 语义分割 全卷积网络 矩形卷积核 融合池化层 条件随机场 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]