Natural Scene character recognition using Markov Random Field

Text recognition in natural scenes is challenging due to the variations on text font, font size, color and scene background. In this paper, we propose a novel method to recognize scene characters based on born-digital template matching, which is formulated by a Morkov Random Field framework by combining both local point features and global spatial structures of scene characters to improve the accuracy. Then the character recognition result of the proposed MRF framework is determined by the votes from multiple one-versus-one classifiers. The experiments on two benchmark datasets show that the effectiveness of the proposed method, which outperforms the recent state-of-the-art natural scene recognition methods.