Exploiting Color Information for Better Scene Text Recognition

This paper presents an approach to text recognition in natural scene images. The main contribution of this paper is the efficient exploitation of colour information for the identification of text regions in the presence of surrounding noise. We propose a pipeline of image processing operations involving the bilateral regression for the identification of characters in the images. A pre-processing step has been proposed to increase the performance of bilateral regression based character identification. The proposed method segments the characters so well that a combination of an off the shelf feature representation and classification technique achieves state-of-the-art character recognition performance. The capability of the framework is further extended for word recognition where a basic string matching criterion is applied to match recognized strings against a lexicon to eliminate errors that occur due to inaccuracies in character classification. We performed extensive experiments to evaluate our method on challenging datasets (Chars74K, ICDAR03, ICDAR11 and SVT) and results show that the proposed method superseded the existing state-of-the-art techniques.