End-to-End Chinese Image Text Recognition with Attention Model

Abstract

This paper presents an attention-based model for end-to-end Chinese image text recognition. The proposed model includes an encoder and a decoder. For each input text image, the encoder part firstly combines deep convolutional layers with bidirectional Recurrent Neural Network to generate an ordered, high-level feature sequence, which could avoid the complicated text segmentation pre-processing. Then in the decoder, a recurrent network with attention mechanism is developed to generate text line output, enabling the model to selectively exploit image features from the encoder correspondingly. The whole segmentation-free model allows end-to-end training within a standard backpropagation algorithm. Extensive experiments demonstrate significant performance improvements comparing to baseline systems. Furthermore, qualitative analysis reveals that the proposed model could learn the alignment between input and output in accordance with the intuition.