Recurrent Highway Networks with Attention Mechanism for Scene Text Recognition

Abstract

Scene Text Recognition is an extremely useful but challenging task and has drawn much attention in recent years. The best of previous model is CNN-LSTM model with attention mechanism, and it can recognize the whole text without character-level segmentation and recognition. Compared with LSTM, Recurrent Highway Networks (RHN), as a popular architecture because of its capability of training deep structure, can preform excellently in plenty of situations and has least parameters. Thus, we employ RHN as decoder and combine attention mechanism with it. Moreover, we integrate feature extraction, feature attention and sequence recognition into an end- to-end framework which can be jointly trained. Our proposed method is conducted on challenging public datasets, such as Street View Text and ICDAR 2003, and outperform the results of the best model in some datasets. Nevertheless, our model only contains 6.3 million parameters that is the minimal size of model for this problem.