模糊和畸变场景图像中的文字识别研究

项目名称： 模糊和畸变场景图像中的文字识别研究

项目编号： No.61471214

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 刘长松

作者单位： 清华大学

项目金额： 82万元

中文摘要： 在文字识别理论和技术研究中，场景图像文字识别是图像文字信息理解的关键技术。在图像模糊和畸变条件下，场景图像中的文字几乎无法识别，后续的图像理解、检索更无从谈起。该问题是进一步发展高性能文字识别技术亟待解决的关键问题。本项目以场景图像文字识别系统为研究平台，深入分析模糊、畸变、缺损对场景图像文字识别性能的影响，重点研究退化场景图像中高性能文字识别理论和方法。具体的研究内容包括模糊图像自动复原、文本图像畸变校正方法、具有模糊和畸变鲁棒性的高性能文字识别方法、缺损字符图像特征提取及识别理论和方法。在以上理论和关键技术研究的基础上，研制成实用的场景图像文字识别系统。该项目的研究涉及图像处理、模式识别、人工智能等多个学科，具有重要的理论研究意义。研究的成果可以推广至基于内容的图像搜索、网络图像信息监控等领域。

中文关键词： 图像识别；模糊图像复原；几何畸变校正；字符提取；文字识别

英文摘要： Among researches of character recognition theory and technology,recognition of scene character image is the key technology for information understanding.When faced with image blurring and distortion,the scene text image almost can not be recognized correctly.It goes without saying the problem of image understanding and indexing based on recognition can not be solved either.Text recognition in degraded scene image is the key problem to be solved for high performance character recognition.This project will be conducted on the benchmark of scene character image recognition system.Based on the analysis of blurring,distorted and broken character images, we will focus on high accuracy character recognition research under these circumstances.The concrete research contents include blurring image restoration,distorted image correction,blurring and distortion invarient character recognition and feature extraction and recognition theory for broken character image theory.On the basis of these research results,a high performance scene character recognition system will be established. This project is significant for theory research since it involves multidisciplinary knowledge, such as image processing,pattern recognition and artificial intelligence.The research results could be extended and applied in content based image indexing,image information monitor on the network etc.

英文关键词： image recognition;blurring image restoration;geometric distortion correction;character extraction;character recognition