Researches On Text Detection Methods In Images With Support Vector Machines And Sparse Representation

Text in image and video plays an important role in describing and annotating image content, and also provides the important clues for video index and retrieval. In general, many high-level applications about image can work well if the text in image would be detected accurately. Accordingly, the text detection in image and video is very important. However, because of the complex background of the images, such as the different type, size, color, inter-space and background textures in characters of the texts in images, the text detection is a critical and challenging research problem in image processing area.In this paper, we first give a short introduction to the characteristics of the text in image, and present a general interview of the current techniques and the development of the text detection as well as the corresponding comparison. Then, two text detection methods are proposed.According to the difference characteristics of texts and backgrounds, we propose a text detection method by combining edge information and support vector machines. In this method, we first obtain the potential text regions by using the edge information and morphology operation. Then, for each potential text region, the wavelet transformation is used to extract co-occurrence feature, and the support vector machines are applied to classify these potential text regions. The results obtained from our experiments demonstrate that the proposed method can automatically detect without human intervention, and perform fast as well as good result.Considering the good performance of sparse representation for image classification, a novel sparse representation based text detection method is presented. In this method, the source signal is respectively decomposed into sparse coefficients over two overcomplete dictionaries, which are designed to reflect the characteristics of text and non-text. Then sparse coefficients are introduced to refine those potential regions, which are achieved by classifying these signals with discriminative sparse representation. Experimental results show that the proposed method can detect text regions from different kinds of complex backgrounds, and has high correct detection probability and robustness. Compared to the classical support vector machines based method, the proposed method performs better in text detection.Finally, the OpenCV (Open Source Computer Vision Library) is used to implement a software system which contains the edge and support vector machine based algorithm as well as the other three classical algorithms in Visual Studio 2008 environment. This system can detect text regions effectively and has high practical value.