【作者】 李朝晖；

【导师】 余英林；

【作者基本信息】 华南理工大学 ， 通信与信息系统， 2004， 博士

【摘要】 本文围绕与MPEG-7相关的关键技术之一——视频文本检测、定位和抽取方法进行了研究，由于视频文本的检测、定位、抽取和识别受背景的复杂度、对比度、分辨率、字体大小、排列方向、字形、运动方式等因素的影响，需考虑的因素较多，算法比较复杂。本文的主要工作即是针对国内外涉足较少的视频文本检测进行了较为深入的研究，为实现如何降低算法的复杂度、提高算法的鲁棒性和准确性这一目标，本文做了如下的工作： 1．提出了一种基于边缘信息和LH的方法检测视频帧文本所在的位置，边缘检测用Sobel算子，将检测到的边缘，经强度、密度检测后，计算其LH的值，经凸台检测后，得到文本区域的定位结果。实验表明这种方法计算简单，能快速地定位文本区域，且不受文本颜色、字形等的限制。对水平或垂直方向排列的视频文本，检测和分割效果良好。 2．提出一种基于小波和LH的检测算法，采用多分辨率的分析方法，不但考虑了水平和垂直方向的高频分量，也考虑了对角方向的高频分量，用具有良好时频局部特性和变尺度特性的小波分析方法提取出不同空间分辨率、不同方向的边缘子图象后经过强度和密度检测，用LH的方法进一步定位文本目标区域。实验表明这种方法由于采用多分辨率的分析方法，可对字体大小差别较大的视频文本进行检测，且不受文本颜色、字形等的限制。对水平或垂直方向排列的视频文本，检测和分割效果良好。 3．提出用形态学进行视频文本检测的两种方法。第一种方法结合了小波多分辨率的分析方法及形态学具有几何形状分析的特点，用小波分析方法提取文本的高频分量后，用形态学的方法进行进一步的处理，这样可不受文本排列方向的限制。第二种方法是提出一种基于形态学的方法提取边缘子图像，所提取的边缘具有良好的形状保持性。在提取到边缘子图像后，再通过区域标记、连通分量分析、形态后处理等一系列步骤对视频帧进行滤波，得到视频文本检测结果，最后再对其进行二值化处理，把文本字符从复杂背景中分离出来。实验表明这种方法可以更多还原

【Abstract】 This dissertation is focus on one of the key techniques related to MPEG-7 standards--the methods of video text detection, locating, and segmentation. By the affectionof the complexity of the background, the contrast of the text with the background, the size of the text, the orientation of the text, the shape of the text and motion method of the text, etc., there are much factors must be considered, so the algorithm is complex. The main purpose of this work is to decrease the complexity of this kind of method, increase its robusticity and accuracy. Several work has been done to achieve this goal:1. A text detection method of video frame based on edge and LH has been proposed. First, the edge image is acquired by Sobel operator, then it’s LH is calculated after intensity and density detection, then the horizontal and vertical convex is detected based LH, finally the text is located. The experimental results demonstrate the simplicity and effectivity of the approach, which is free of the restriction of the text color, the text shape, and etc. This method works well in the horizontal and vertical text.2. A text detection method of video frame based on wavelet and LH has been proposed. This method, which based multiscale analysis, not only take horizontal and vertical high frequency component into consideration, but also considered the diagonal high frequency component. First, the edge image is acquired by Wavelet analysis which has fine temporal-frequency local feature and multiscale feature, then it’s LH is calculated after intensity and density detection, then the horizontal and vertical convex is detected based LH, finally the text is located. By adopting multiscale method, it can detect variable font size. The experimental results demonstrate the simplicity and effectivity of the approach, which is free of the restriction of the text color, the text shape, and etc. This method also works well in the horizontal and vertical text.3. Two approaches based on morphology to detect video text are proposed. In the first approach, the high frequency component of the text is first acquired by wavelet更多还原