基于树修剪和多特征融合的场景文本检测  

Scene-Text Detection Based on Tree Pruning and Multi-Cues Integration

作　　者：肖诚求 吉立新[1] 高超[1] 李邵梅[1] 

机构地区：[1]国家数字交换系统工程技术研究中心,河南郑州450002

出　　处：《信息工程大学学报》2015年第5期590-595,601,共7页Journal of Information Engineering University

基　　金：2014年国家科技支撑计划项目(2014BAH30B01)

摘　　要：为了解决最大稳定极值区(MSER)提取过程中产生的大量重复文本区域和非文本区域难以被剔除影响算法精度的问题,提出了一种基于树修剪和多特征融合的场景文本检测方法。首先提取出边缘叠加的MSER作为文本候选区域;其次设计了一种MSER树修剪算法剔除重复文本区域;然后采用贝叶斯分类器融合多特征剔除非文本区域;最后设定了一系列相似性标准合并文本区域。ICDAR 2011数据集(f=76.8%)上的实验结果低于目前最好的算法[19],但算法在速度上具有明显的优势。Repeat components and non-characters caused in extracting maximally Stable Extremal Regions （MSER） are problematic for algorithm accuracy. A scene-text detection method based on tree pruning and multi-cues integration is proposed to solve this problem. Firstly, edge-sharped MSERs are extracted as character candidates. Secondly, a novel MSER tree pruning algorithm based on regularized variations is designed to filter the repeat components. Thirdly, a Bayesian classifier is exploited to integrate multi-cues to filter non-character candidates. Finally, a series of similarity cri- teria are designed by grouping characters into words. Experimental results on ICDAR 2011 dataset （f=76.8%） show that the performance of this method is a little lower than that of the state-of-the- art algorithm performance, but the speed is significantly faster.

关 键 词：MSER树修剪 多特征融合 场景文本 最大稳定极值区 贝叶斯多特征融合分类器 

分 类 号：TP242.62[自动化与计算机技术—检测技术与自动化装置]