一种针对汉字特点的场景图像中文文本定位算法  

Chinese Scene Text Localization Algorithm Based on Character Features

作　　者：张伟伟[1] 汤光明[1] 孙怡峰[1] 苏伟 

机构地区：[1]信息工程大学 [2]68310部队

出　　处：《信息工程大学学报》2014年第6期729-736,共8页Journal of Information Engineering University

基　　金：国家自然科学基金资助项目(61101112);河南省科技攻关基金资助项目(122102210047)

摘　　要：针对场景图像中的中文文本定位问题,提出一种基于汉字特点的中文文本定位算法。算法首先对图像提取MSER(最大稳定极值区域),然后通过剪枝策略对存在嵌套关系的MSER进行取舍,得到候选笔画区域;计算候选区域的笔画宽度,作为闭操作的结构元参数并对图像进行动态闭操作,以消除同一汉字笔画之间的间隙,得到候选汉字区域;利用结构和角点规则过滤掉非汉字区域,并用颜色规则聚类得到候选文本区域;最后基于图像文本大都横向或纵向排列的规律,通过分析各组连通区的分布情况,对候选文本区域进行确定。在测试数据集上的实验表明,算法对于汉字与背景色差明显、汉字成行或成列排列等场景文本具有定位稳定性和准确性。The paper proposes a kind of location algorithm based on the feature of Chinese charac- ters, for the problem of locating Chinese text in natural scene images. The method extracts the maxi- mum stable extremal regions（MSERs）from the image first. Then the pruning strategy is performed to make trade-offs between the MSERs with nested relation, getting the candidate stroke areas of high recall ratio. In addition, the method calculates the stroke width, and treats it as the parameter of structural element in the close operation. Then, the dynamic close operation is used to eliminate the interval of different strokes in the same character, getting the candidate character areas. Based on the above operation, the non-character areas are filtered out by the structure and angular point rules, and the candidate areas are grouped by color clustering model. In that way, the candidate text areas are extracted. As the texts always obey the rule of arranging in rows or lines, the position attribution of each group is analyzed to validate the text areas. The experiment on testing dataset shows that the proposed algorithm is more stable and accurate in localization for the scene texts, which differ from the background in color and are arranged in rows or lines.

关 键 词：闭操作 文本定位 颜色聚类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]