Support vector machines for text location in news video images

The aim of this paper is to show the applicability of support vector machines (SVMs) for the problem of text location and to propose an SVM-based method for locating texts in news video images. The proposed method is based on observations that texts in digital video have distinct textural properties that can be used to discriminate texts from the background and an SVM can be trained to be a texture classifier. An SVM is used for classifying a pixel into text or non-text by analyzing the textural properties of video image. To achieve multi-scale location, the video image is incrementally resized and the location process is performed over each of these resized images.