电子学报 ›› 2018, Vol. 46 ›› Issue (6): 1436-1444.DOI: 10.3969/j.issn.0372-2112.2018.06.024

• 学术论文 • 上一篇    下一篇

基于自适应色彩聚类和上下文信息的自然场景文本检测

邹北骥1,2, 郭建京1,2, 朱承璋1,2, 杨文君1,2, 徐子雯1,2   

  1. 1. 中南大学信息科学与工程学院, 湖南长沙 410083;
    2. 中南大学眼科医学影像处理研究中心, 湖南长沙 410083
  • 收稿日期:2016-12-22 修回日期:2017-02-24 出版日期:2018-06-25 发布日期:2018-06-25
  • 通讯作者: 朱承璋
  • 作者简介:邹北骥,男,1961年生于湖南邵阳.中南大学教授,博士生导师,研究方向为图像处理与计算机视觉.;郭建京,男,1990年生于湖南祁阳.硕士研究生,研究方向为计算机视觉与模式识别
  • 基金资助:
    国家自然科学基金(No.61573380,No.61702559);湖南省科技计划项目(No.2017WK2074);中南大学创新创业师生共创项目(No.2017gczd016)

Natural Scene Text Detection Based on Adaptive Color Clustering and Context Information

ZOU Bei-ji1,2, GUO Jian-jing1,2, ZHU Cheng-zhang1,2, YANG Wen-jun1,2, XU Zi-wen1,2   

  1. 1. School of Information Science and Engineering, Central South University, Changsha, Hunan 410083, China;
    2. Center for Ophthalmic Imaging Research, Central South University, Changsha, Hunan 410083, China
  • Received:2016-12-22 Revised:2017-02-24 Online:2018-06-25 Published:2018-06-25

摘要: 自然场景文本检测是图像内容分析和理解的重要前提.本文提出一种基于自适应色彩聚类和上下文信息分析的方法,用于检测自然场景图像文本.首先,将层次聚类和参数自学习策略结合,设计一种自适应色彩聚类方法,提取图像中的候选字符.该自适应色彩聚类方法能针对不同图像自动学习权重阈值,有较好的字符召回率.然后,利用文本中字符成行出现的性质,设计一种基于上下文信息的字符验证策略,既能保证较高字符召回率,也能有效移除非文本字符.最后,合并字符构建文本行,并通过后处理得到文本检测结果.在ICDAR2013公共数据集上的实验结果表明:本文分别获得74.17%的召回率,83.40%的准确率和78.52%的F得分.与其他文本检测方法相比,本文获得了较好的文本检测性能,说明本文方法的优越性.

关键词: 自然场景文本检测, 自适应色彩聚类, 上下文信息, 自学习策略

Abstract: Natural scene text detection is an important task for image analysis and understanding.In this paper,a natural scene text detection method is proposed,using adaptive color clustering and context information analysis.Firstly,combining hierarchical clustering with self-learning strategy,we design an adaptive color clustering method,which learns clustering weights automatically and generates high character recall.Then,considering text in images usually containing several characters,we propose a character verification strategy based on image context information,which can guarantee high character recall and remove non-text components at the same time.Finally,characters are merged to text lines,and further post-processing is applied to generate final text detection results.Experiments on the ICDAR2013 publicly available dataset show that we obtain recall of 74.17%,precision of 83.40% and F-score of 78.52%.Compared with other text detection methods,our method obtains better text detection performance,indicating superiority of the proposed method.

Key words: natural scene text detection, adaptive color clustering, context information, self-learning strategy

中图分类号: