Research On Scene Text Recognition Based On Deep Leraning

Scene text recognition is a universal text recognition technology.In recent years,it has become a key research direction in the computer field.Traditional Optical Character Recognition(OCR)has achieved high recognition performance in document images,but it is difficult to deal with the complexity of scene text in terms of font,distribution and background.In Recent years,deep learning technology has developed rapidly and played a leading role in the field of OCR.Therefore,based on the deep learning method,this paper focuses on the problem of scene text recognition.The main research contents are as follows:(1)A text image rectification method is proposed.Aiming at the problem of text line tilt,a tilt rectification method based on connected region analysis is proposed.At first,the method obtains the connected region of the input image,and then filters the non-character connected region according to the prior knowledge of the scale and size of the character,and finally estimates the tilt angle of the text row by calculating the tilt angle between the connected regions.The experimental results show that the proposed method can effectively improve the recognition performance of the same framework.(2)Design a sequence-to-sequence text recognition network.Firstly,based on the Convolutional Recurrent Neural Networks(CRNN)network,the proposed image rectification method is used to rectify the input image to reduce the complexity of the problem.Secondly,the convolution part of the CRNN is added to the BN layer.While improving the convergence speed of the convolutional neural network andreducing the sensitivity of the network to the initialization weight,it also plays a regularization role and improves the system performance;Thirdly,the long short-term memory network used in the CRNN,that is replaced with the gated recurrent unit to reduce the complexity of the network.Experiments show that compared with the original framework,the proposed framework reduces the network time and space complexity,and can effectively improve the recognition performance of scene text.(3)A sequence-to-sequence scene text recognition method based on attention mechanism is designed.To begin with,for the shortcomings of the connectionist temporal classification method,in the sequence decoding stage,the connectionist temporal classification is replaced by the attention mechanism based decoder.Then because of the shortage of network feature extraction,the feature extraction network is replaced with a ResNet network in this framework.Finally,under the improved framework,verifying the text line rectification method proposed in this paper can improve the recognition performance of the system.The experimental results show that it achieves a better result in the COCO-Text dataset.