A feature learning method for scene text recognition

Abstract

Reading text in scene images is a challenging task and is still an active research nowadays. The difficulties come from low resolution, complex background, non uniform lightning or blurring effects of scene images. This paper focuses on recognizing characters in scene images based on the feature learning method proposed in [6] and the conclusion on comparison between sparse coding and vector quantization in [8] to build better feature representations before training the model by using SVM. We asset the performance of the proposed method on some popular scene image datasets such as ICDAR 2003 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> and Chars74k <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">4</sup> . Experimental results show that our proposed system has reached an encouraging recognition rate for both ICDAR 2003 and Chars74k datasets. More specially, our system archived 83.8% (62-class problem), 87% (36-class problem) of recognition rate on ICDAR 2003 Sample subset (698 images), and 73.8% accuracy on GoodImg subset (7705 images) of Chars74K dataset. In this work, our contribution is that we applied the ideas as well as the conclusions in [8] for scene text recognition problem and the experimental results show that our system outperforms other existing methods.