结合笔画方向信息与SVM的英文文字检测  

English text detection combining stroke direction information with SVM

作　　者：黄为[1] 王云宽[1] 李兵[1] 吴少泓[1] 

机构地区：[1]中国科学院自动化研究所,北京100190

出　　处：《中国图象图形学报》2011年第6期989-994,共6页Journal of Image and Graphics

基　　金：中国科学院自动化研究所科技产业发展基金项目(DC07J03)

摘　　要：文字检测是文字信息提取系统中最重要的环节。针对在非均匀光照或背景图案复杂等恶劣环境下文本行难以被检测的难题,设计一种通用的基于笔画方向信息图的通用文字检测算法,该算法采用由粗到细的实现框架。在粗的文字行定位中,利用Haar小波和LBP描述符,建立与原图像相对应的笔画方向信息图,经滤波,连通域分析以及PPA后得到候选的文字行;在文字行精确分类中,利用SVM分类器,并结合多种文字的纹理特征,确认最终的文本行区域。针对图片数据库ICDAR03的实验表明该算法能在不同条件下快速,准确地检测出文字区域,文字检测的精确率为0.64,召回率为0.67。Text detection plays an important role in the system of text information extraction. To solve the problem of detecting text in the bad settings such as complicated Text-background, uneven illumination and so on, a generation text detection method based on stroke direction map is proposed, which utilizes a coarse-to-fine framework. In the coarse step, it firstly builds the stroke direction Information map of the source images with Haar wavelet and LBP descriptor. Then the candidate text lines can be obtained through filtering, component analysis and projection profile analysis of the direction Information map. In the final step, it makes use of SVM classifier with kinds of text texture features to verify the real text lines. Experiment results on ICDAR03 show that the proposed algorithm can quickly and accurately detect texts region in different settings.

关 键 词：文字检测 HAAR小波 LBP 支持向量机 

分 类 号：TP391.12[自动化与计算机技术—计算机应用技术]