Design And Implementation Of Chinese Character Recognition Model Based On Deep Neural Network

Chinese character recognition refers to the recognition of Chinese character information in pictures and predicting real sequence of words in pictures by extracting the character information features.It is one of the popular research areas in the field of computer vision,it can be used in a wide range of applications.Based on the deep learning method,this thesis studies words recognition deep learnning algorithm model.At present,the classical CRNN(Convolutional Recurrent Neural Network)model achieves a high accuracy in the field of word recognition,but it can be still improved due to the influence of different text sizes,low image resolution and easily confused similar words.Combined with the Inception network structure and attention mechanism,this thesis improves the architecture of CRNN model.The main research work of this thesis is showed as follows:(1)A character recognition algorithm based on Inception network is proposed in this thesis.The principle,structure and performance of Inception network were studied and analyzed.Ineption network were compared with several commonly used deep networks,such as CNN,VGG and Res Net to prove the effectiveness of Inception network structure.Based on the Inception network,a text recognition model was designed.Inception network module was added in the feature extraction layer,the multi-filter branch of the Inception network module can learn local regional features in different scales,and the network features of the same level of multi-kernel convolution enhances the ability of feature extraction.According experimental results,the accuracy of the new character recognition model was averagely improved by 5.97% compared with that of the benchmark model on the ICDAR17,ICDAR19 and self-made dataset.(2)A character recognition algorithm based on attention mechanism is proposed.In order to solve the performance problem of BLSTM recurrent layer in CRNN model,an attention mechanism and LSTM combied structure was designed to replace the BLSTM network structure in the benchmark model,which enhanced the capability of recurrent layer auxiliary inference and identification.The attention mechanism calculates different feature contribution degree to generate weight probability distribution.The recurrent unit selects and recombines the network sequence according to the weight value,and the output of predicted sequence which is closer to the real label.According experimental results,the accuracy of new character recognition model improved by 5.11%,6.76% and 10.09% averagely under three evaluation criteria of line text accuracy,longset common substring accuracy and editing distance accuracy compared with the benchmark model on the ICDAR17,ICDAR19 and self-made dataset.(3)A small character area detection and recognition system is designed.This thesis adopts EAST text area detection model and the new character recognition model proposed in this thesis to design character area detection and recognition system combining relevant image preprocessing and data acquisition technologies.