Scene Text Recognition by Attention Network with Gated Embedding

Abstract

Recurrent attention based encoder-decoder model is one of the most popular frameworks for scene text recognition. However, most methods in this category only use standard recurrent attention network as the decoder. In this paper, in order to alleviate the problem that standard attention network relies on the previous output character overmuch, we propose an attention network with gated embedding for scene text recognition. The proposed attention network with gated embedding (GEAN) adopts a gated embedding to adaptively reset the input information from the embedding vector of previous output character for recurrent attention network. The gated embedding is constructed by adding an adaptive embedding gate based on the degree of correlation between the hidden state vector and the embedding vector of the corresponding character at the same time step. We verify the effectiveness of GEAN for scene text recognition through extensive experiments on both regular and irregular scene text datasets. The performance of GEAN is shown to be superior to the standard recurrent attention based decoder and is comparable compared with state-of-the-art methods.