Text recognition in natural scenes based on deep learning

Aiming at the problems of character segmentation and dictionary dependence in text recognition in natural scenes, a text recognition algorithm based on Attention mechanism and connection time classification (CTC) loss is proposed. Convolutional neural network and bidirectional long short – term memory network are used to realize image feature coding, which avoids the gradient vanishing problem of recurrent neural network (RNN) with the increase of time. And the Attention-CTC structure is used to decode the feature sequence, which effectively solves the problem of unconstrained attention decoding. The algorithm avoids extra processing of alignment and subsequent syntax processing, and improves the speed of training convergence and significantly improves the recognition rate of text. It has a certain research value in recognition accuracy. Experimental results show that the algorithm has good robustness to text images with fuzzy fonts and complex background.