2D Attention Scheme Based Irregular Scene Text Recognizer

Irregular scene text,a difficult subproblem of Optical Character Recognition(OCR),which has complex layout in 2D space,is challenging to most previous scene text recognizers.Recently,all methods in academia and industry will be splited into three categories.Some irregular scene text recognizers rectify the irregular text to regular text image with approximate 1D layout via thin plate splines(TPS).Some transform the 2D image feature map to 1D feature sequence,and others cite the notion of connectionist temporal classification(CTC),which is from natural language processing.Though these methods have achieved good performance,the robustness and accuracy are still limited due to the loss of spatial information in the process of 2D to 1D transformation.In this paper,we will predict characters of 2D layout irregular scene text directly via 2D attention mechanism.In this paper,we propose three modules to recnogize irregular scene text,which are 2D feature extraction module,relation attention module and parallel attention module,respetively.Different from all of previous,we propose a framework which transforms the irregular text with 2D layout to character sequence directly via 2D attentional scheme.To address the problem of irregular scene text recognizing.In this paper,three moduls are proposed to deal with the information loss due to the process of 2D to 1D transformation.First,2D feature extraction module is proposed.This module is used to extract high-level feature of 2D image.And to avoid information loss,this module modifies the backbone of powerful residual network.Wath's more,this paper utilizes a relation attention module to capture the dependencies of feature maps and a parallel attention module to weight the output of relation attention module,and parrallel attention module will give attentional weighted output to decoder,which will decode all characters in parallel.To sum up,parallel module will make our method more effective and efficient.Extensive experiments on several public benchmarks as well as our collected multi-line text dataset show that our approach is effective to recognize regular and irregular scene text and outperforms previous methods both in accuracy and speed.In terms of speed,our proposed method is 2.1 times faster than other algorithms proposed by academia.As for accuracy,ours is 7.3% higher than others in terms of irregular scene text dataset.