Research On Optical Character Recognition Technology Based On Deep Learning

Optical character recognition(OCR)technology is now mainly used in document and certificate recognition.Document recognition can digitize the text information of handwritten document or printed document,extract the effective information quickly and accurately.The rapid development of deep learning applied to OCR at this stage can not only be applied to different scenarios,but also the accuracy of character recognition can be improved.The main work of this paper is to build a handwritten character recognition framework by using deep learning,which is implemented in smartphone.The steps of OCR are mainly text region detection,character cutting and recognition.Character recognition is the research focus of this article,so it is optimized in combination with deep learning.In this thesis,we first summarize the deep learning network which can be used in character recognition.The deep learning framework,convolutional neural network,and the recurrent neural network framework,Tensorflow,are analyzed in detail.Secondly,because the size of the dataset is critical to training the model,the data set is augmented.Based on the variability of the text line image writing scale,multi-scale text is used to expand the data set.Then,improve feature extraction that greatly affects the performance of the system.Use the residual idea to improve and optimize the convolutional neural network appropriately,change the method of information dissemination during feature extraction,and improve the accuracy of feature extraction.Finally,to address the problem of poor portability of the device during traditional OCR applications,improvements were made.In the past,when OCR technology was used to recognize text,the document image was mainly obtained through a device such as a scanner or a camera,and the applicable situation was limited.Therefore,the research of handwritten character recognition technology based on smart phones is proposed,and the computing power of smart phones has been greatly improved in recent years to achieve document image acquisition and character recognition on smart phones.In this paper,the effectiveness of the residual-convolutional network framework for document image feature extraction is verified by experiments,and the character recognition function of handwritten document image implemented on Android smartphone increases the practicability.