Natural scene text recognition using convolutional recurrent neural network

In this article, we explore the scene text recognition problem, which is one of the challenging sub-fields of computer vision. Recently, deep learning has achived state-of-the-art performance for recognition task. The convolutional recurrent neural network (CRNN) architecture is explored for this task, which consists of feature extraction, sequence modeling. Moreover, an attention mechanism is introduced in our study. Unlike many of previous scene text recognition systems, the proposed architecture has several advantages: the model can be trained using the end-to-end manner and the CRNN can deal with the sequences of arbitrary length. Comparing the detection results of several mainstream CNN network structures, the experimental results show that the accuracy of the detection results is improved, and false positives are reduced, which clearly demonstrate its effectiveness.