【作者】 徐莉莉；

【导师】 唐降龙；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2006， 硕士

【摘要】 图像和视频中的文本字符,是图像高层语义内容的一个重要来源,它包含许多非常重要的有用的信息,如街道名称、商店名称、路标、交通标示、字幕等,这些信息对于图像和视频资料的自动注释、索引、压缩等方面有重要的参考作用。通常情况下根据文本对象的存在形式将文本分为人工文本和场景文本。目前的研究主要集中于对人工文本的研究,而场景文本的研究刚刚起步。文本提取的研究思路主要集中于应用各种图像处理的方法将文字区域的边角特征、色彩特征和纹理特征突出,然后根据一定的分类算法来划分候选的区域,经过进一步的后处理来最终确定实际的文本区域。本文针对文本区域提取这个问题来进行研究,包含预处理、多分辨分析、特征提取、分类(检测)、区域提取五个步骤来解决文本区域的准确提取问题。由于图像上字符的大小变化范围很大,为了能够找出大小不一的文本区域,在多分辨分析中,本文将图像分解为原分辨率的1/1、1/2和1/4的三幅子图,然后对每幅子图应用相同的文本区域定位算法。在特征提取步骤中,本文把一致性H应用到文本区域提取领域,使用边缘空间映射和一致性H空间映射两种方法得到特征图像,并比较了两种空间对于文本提取的影响;对得到特征图像,使用滑动窗口比较了提取不同维数的纹理特征作为特征向量的结果。在分类检测时,通过对比支持向量机(Support Vector Machine SVM)和BP前馈网络的分类效果,选择使用支持向量机作为分类器。根据对多幅子图的分类结果进行图像融合来生成候选区域,然后逐步对候选区域进行基于密度的区域确认和平滑后处理等步骤确定最终的文本区域。通过实验表明使用一致性H变换后提取25维特征,经过支持向量机分类可以较为准确的提取出文本区域,得到较为理想的评价指标。更多还原

【Abstract】 Text embedded in picture and video is an important resource for picture’s high level content understanding, it contains some useful information, for instance street names, shop names, signposts, traffic signals, titles, etc. That information is important in indexing, compressing of the picture and video.In general, it can be divided into artificial text and scene text according to the existent form of text, present research mainly concentrates on artificial text, and the research of the scene text has just started. The mainly method is using different image processing methods to enhance character’s corner, color and texture features, then classifying text and background to form the candidate regions, then rectifying the real text area through the postprocess.In this thesis, we study on text detection. It includes five parts: pre-process, multi-scale analysis, feature extraction, classification and text area extraction. The size of text in scene image varies a lot, to extract different size text, in multi-scale process, we make the image into three resolutions: 1/1, 1/2, 1/4. In feature extraction step, we apply homogeneity into text detection, and we compare using the gradient, edge extract and homogeneity mapping to enhance corners and texture features, and then use a slip window to get different kinds of texture features as the feature vector, and then after comparing the accuracy result of SVM and BP neural network, we choose SVM as the classifier. We get the candidate area by integrating the SVM results of different scales, and use density-based candidate area confirmation and image smoothing postprocess to get the final text area.Through experiment, the statistics in Homogeneity area play well in feature extract and SVM can be a good classifier, so using this algorithm can reach an ideal result.更多还原