面向敏感网络图像过滤的原生数字图像文本提取关键技术研究

项目名称： 面向敏感网络图像过滤的原生数字图像文本提取关键技术研究

项目编号： No.61201424

项目类型： 青年科学基金项目

立项/批准年度： 2013

项目学科： 电子学与信息系统

项目作者： 王恺

作者单位： 南开大学

项目金额： 24万元

中文摘要： 原生数字图像是用计算机软件生成的图像。不良文本信息可以很容易地转为原生数字图像在网络传播，从而避过文本过滤技术。原生数字图像文本提取工作面临分辨率低、边缘柔化所引起的连通体生成困难这一问题，使得以往复杂图像文本提取研究中常用的基于连通体分析的方法失效。本课题拟构建多级优化模型，通过像素点级、连通体级、文字级的同步优化解决连通体生成困难问题；采用基于样本库的自适应学习解决优化模型的参数选取问题；利用彩色图像过分割等方法解决模型求解效率问题，并基于过分割结果融合及软决策方式解决超像素边界与文字边界的匹配问题。针对敏感网络图像过滤应用，将敏感关键词信息融入到优化模型中，以高召回率检测敏感图像，解决不良文本内容以图像方式在网络传播所带来的社会问题。

中文关键词： 文字信息；原生数字图像；文本分割；光学字符识别；基于内容的图像过滤

英文摘要： Born-digital images are generated directly on the computer. Harmful text transmitted online is easily converted into born-digital images to avoid text-based filtering. In previous studies on text extraction from complex images, component connector-based methods have been widely used. However, it is a difficult task for born-digital images to generate component connectors because born-digital images are inherently low-resolution for online transfer and often suffer from anti-aliasing. To solve this problem, a hierarchical optimization model, that allows integration of features computed at different levels of the quantization hierarchy, will be set up in this study. Parameters of the model will be automatically estimated by supervised training. To reduce the computational time for solving the hierarchical model, superpixels, that generated by over-segmentation of color images, will be considered as an alternative for pixels. Furthermore, techniques of using multiple segmentations or soft decision will be applied during over-segmentation to ensure that superpixels align well with the boundary of characters. Aiming at the application of content-based web image filtering, predefined keywords will be integrated into the hierarchical model, so that the images embedded with harmful text can be detected with high recall

英文关键词： text information；born-digital image；text segmentation；character recognition；content-based image filtering