Abstract:

In order to solve the problems of the difficulty of character segmentation in text recognition and the recognition accuracy dependent on dictionary in natural scene, a text recognition algorithm with attention mechanism and connection time classification loss was proposed. The convolutional neural network and the bidirectional long-term memory network were used to extra the feature of the image. The Attention-CTC structure was used to decode the feature sequence, which effectively solved the problem of Attention decoding unconstrained. The algorithm avoided additional alignment preprocessing and subsequent syntax processing on the tag, which sped up the training convergence rate and significantly improved the text recognition rate. Experimental results showed that the algorithm was robust to texts with complex fonts and complex backgrounds.

Key words: text recognition, connection time classification, convolution neural network, recurrent neural network, multi-scale feature extraction, attention mechanism