自然场景图像文字检测研究综述  

Text detection from natural scene image: a survey

作　　者：郭芬红[1] 谢立艳 熊昌镇 GUO Fenhong;XIE Liyan;XIONG Changzhen(College of Sciences,North China University of Technology,Beijing 100144,China;Beijing Key Laboratory of Urban Intelligent Control,Beijing 100144,China)

机构地区：[1]北方工业大学理学院,北京100144 [2]城市道路交通智能控制技术北京市重点实验室,北京100144

出　　处：《计算机应用》2018年第A01期173-178,共6页journal of Computer Applications

基　　金：国家十三五重点研发计划项目(2016YFB1200402)

摘　　要：文字是图像内容的重要表达,随着基于内容的图像检索技术的发展,复杂场景图像下的文字检测技术越来越受关注,针对此类图像对现有的主流算法进行了详细的研究。文字检测算法主要包括候选文本区域提取和文本/非文本分类两大核心步骤。首先,总结了近5年的21种主流算法在公开数据集ICDAR上的文字检测效果,数据显示现有文字检测算法依然存在低召回率的问题,召回率最高为0. 83;其次,对候选文本区域提取和文本/非文本分类两大核心步骤中典型算法的优缺点及存在的问题进行了详细的分析;最后,探讨了文字检测未来的发展趋势,并提出了4种可能的研究方向。Text is an important expression of image content. With the development of content-based image retrieval technology, more and more attention has been paid to the algorithm of text detection from natural scene images. For natural scene images, the mainstream algorithms were studied in detail. There are two steps of text detection： extracting candidate regions and classifying local regions. Firstly, the experimental results on the open data set ICDAR of mainstream text detection algorithms were summarized, and the data shows the existing algorithms have low recall rates and the highest recall rate is only 0.83. Secondly, the advantages and disadvantages in the two steps of the mainstream algorithm were discussed, and the problems of these algorithms were analyzed in detail; Finally, the development trend of text detection was discussed and four research areas were proposed.

关 键 词：文本检测 场景图像 最大稳定极值区域 关键点检测 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]