Text Detection using Coarse detection and SVM Classification

Abstract

In this paper a new approach to detect text in natural images is described using different detection methods. The end result will be that the text will be segmented from the image and can be used for different purposes. The approach is split up in two parts, a coarse detection step to extract patches from the image and a fine detection step that uses feature descriptors and a support vector machine in order to increase the precision of the coarse detection step. The methods used for the coarse detection are global threshold, mean threshold, Gaussian threshold, local binary pattern, maximum gradient difference filter and maximum difference filter. These methods are compared and the best results are used in combination with the fine detection. The feature descriptors used in the fine detection are Histogram of Oriented Gradients, Co-occurrence histogram of orientated gradients and local binary patterns. In order to increase the quality of the coarse detection a projection step is used. The approach performs on precision level worse than the current state-of-the-art methods, but has a better recall rate than most methods.