【作者】 于佳；

【导师】 王燕；

【作者基本信息】 哈尔滨工程大学 ， 计算机软件与理论， 2010， 硕士

【摘要】 视频中的人工文本是表达视频语义信息的重要手段之一,对它的有效检测、定位与提取对于视频的分析理解和检索等应用都具有十分重要的意义。论文在对现有文本检测方法的研究基础上,针对大多数文本检测方法受到文本位置与语种的限制问题,给出一种视频中非限定性人工文本区域的检测方法。首先,重点研究SOM神经网络的特点,提出一种基于有导师学习的SOM神经网络分类器,从而使SOM在文本检测中能够对分类后的像素进行分类标记。其次,根据视频中人工文本区域存在强烈边缘的特点,采用Sobel边缘检测算子提取图像的边缘特征,提取出的特征作为特征向量的一部分。因为视频中人工文本区域与背景存在着明显的边缘,而Sobel算子不但具有方向性并且它检测到的边缘一般大于两个像素,能较好地描述图像边缘特征。然后,针对人工文本区域出现位置不同的问题,结合分类器的结构,给出区分人工文本区域与背景的特征向量的基本形式,利用有导师SOM分类器初步定位目标文本区域。最后,采用基于形态学的方法对目标文本区域进行处理,用以处理由于分类器的误判而出现的一些噪声点和空洞,并对精确定位的目标文本区域进行后期处理。实验表明,论文给出的方法能够有效检测并定位视频中的非限定性文本区域。更多还原

【Abstract】 Artificial text in video is an important solution of expressing the semantic information of video. The effective detection, location and extraction of artificial text have very great significance for the understanding and retrieval of video and other applications.By the depth analysis of the existing text detection method in which most method has the text location and language constraints, a video unlimited artificial text area detection algorithm has been proposed in this thesis. Firstly, by analyzing, the features of SOM, SOM neural network classifier based on supervised learning is proposed, so that SOM neural network is able to marked the pixels classified. Then, by getting detail knowledge of edge feature of the video artificial text area, Sobel edge detection operator is the effective way to extract the image edge features. The features abstracted were used to be a part of feature vector. Because there’s a clear edge between video artificial text area and background, while the Sobel operator not only has the direction but also the edge it gets in general is greater than 2 pixels. So it can be used to describe the image edge features better. For the problem of different occurrences of artificial text area, combined structure of the classifier proposed, a basic form of feature vectors is given. Then the supervised SOM classifier is used to classify them to get the initial target text area. Finally, the morphology-based approach is used to process the target text area to remove the noise points to eliminate the probability of misclassifying of the classifier. Then perform the post-processing of the precise positioning target text area. Experiments show that the method proposed can effectively detect and locate the video unlimited artificial text area.更多还原