Text detection and recognition in natural scene images

Abstract

Text detection and recognition in natural scene images plays an important role in content analysis of images. In this paper, based on the characteristics of scene text, we propose a robust text detection and recognition method using Maximally Stable Extremal Regions (MSER) and Support Vector Machine (SVM). Different from the end to end text recognition, we split the recognition problem into detection and recognition procedure. Firstly, in the detection stage, in order to extract potential text as much as possible, we use MSER and color clustering to extract connected component. Then, for the obtained candidate connected component, we use visual saliency and some prior information to filter non-text regions. Finally, we can obtain word image by text line generation. In the recognition stage, we use vertical projection to segment word images, then recognize character in SVM based framework. The experiment results evaluated on standard dataset show that with a small amount of prior information and simple segment strategy, the proposed method has a better performance compared to conventional text detection and recognition method.