一种基于图像矩和纹理特征的自然场景文本检测算法  

Moment and Texture Based Algorithm for Text Detection in Natural Scene Images

作　　者：杨玲玲[1] 叶东毅[1] 

机构地区：[1]福州大学数学与计算机科学学院,福州350100

出　　处：《小型微型计算机系统》2016年第6期1313-1317,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61473089)资助

摘　　要：文本检测是许多文本识别应用的必要前提,有效地在自然场景图像中检测定位文本能大大提高文本识别的效率.针对现有研究中存在文本检测率不高的问题,利用字母与非字母在Hu矩特征上的差异性和文本与非文本在纹理特征上的差异性,提出了一种基于矩和纹理特征的自然场景文本检测算法.该算法首先通过提取最大稳定极值区域(MSER)找出自然场景图像中存在的候选字母;其次,为了有效地删除非字母候选对象,算法在字母分类器中引入Hu矩特征刻画候选字母的几何特征;接下来算法利用自然场景图像中文本具有相似性的特征,通过单链接聚类得到候选文本;最后针对文本和非文本候选的纹理差异,在文本分类器中引入共生纹理特征以删除非文本候选.实验结果表明,与同类算法相比,该算法在召回率和f_measure值上有较大的提高,因此是一种有效的检测方法.Text detection is a necessary prerequisite for many text recognition applications,effective detection and location of text in natural scene images can greatly improve the efficiency of text recognition. In view of the problem of low detection rate in current study ,we propose a moment and texture based text detection algorithm making use of the difference in Hu-moment between characters and noncharacters and that in texture between texts and non-texts. The algorithm firstly finds candidate characters by maximum stable extremal regions （MSER} algorithm ;as Hu-moment features are rotation,zoom and translation invariant, the algorithm uses a character classifier incorporated with Hu moment features to delete non-character candidates;after that, as characters in natural scene images tend to have similar features, the algorithm adopts the single-link clustering algorithm to get text candidates;finally, as text and non-text candidates vary greatly in texture,a text classifier mainly trained with symbiotic moments is used to delete non-text candidates. Compared with similar algorithms,the proposed algorithm outperforms in recall rate and f-measure value. Therefore,the algorithm is an effective one.

关 键 词：文本检测 MSER 单链接聚类 HU矩 共生纹理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]