Research On Attention-based Character Recognition

As the carrier of information inheritance in human society,characters widely exist in our daily life and play a very important role in scene understanding.With the rapid developments of computer technologies,more and more new application scenarios need the support of character recognition technology.The development of deep learning has brought many new technologies for character recognition,among which the attention mechanism is a brand new pipeline.Based on deep learning and attention mechanism,this paper proposes new solutions to some hot issues of character recognition and text recognition.The main contributions of this paper can be summarized as follows:1.An exploration of the application of attention mechanism to the task of character recognition: New method for few-shot offline handwritten Chinese character recognition.This paper explores the drawbacks of previous works in solving this problem and proposes radical aggregation network.This method introduces printed Chinese characters as support samples to solve the problem of missing structural information of unknown classes of handwritten Chinese characters.Besides,a character analysis decoder is proposed to output the categories of unknown handwritten Chinese characters in end-to-end manner.Through experiments,this paper verifies that the radical aggregation network has a very good performance in few-shot handwritten Chinese characters recognition task,and at the same time maintains a high performance in conventional handwritten Chinese character recognition.2.Optimized attention mechanism for text recognition: Decoupled attention network for text recognition.To solve the problem that the attention mechanism is prone to misalignment on long text,this paper observes that the cause of this problem is the coupled alignment and decoding process.Therefore,this paper proposes a decoupled attention network,which decouples the alignment process from the decoding process so that the alignment process will not be affected by the decoding results.Experiments show that the decoupled attention network can effectively solve the misalignment problem in long text recognition,so its performance is excellent.Besides,it has also achieved good results in the scene txt recognition tasks.In summary,this paper explores the application scenarios and modeling methods of attention mechanism for character recognition.I hope this work can inspire the researchers and engineers in related fields!...