The Research Of Low Quality Text Recognition Based On Deep Learning

Optical character recognition is a very important field in pattern recognition which has been widely concerned in recent decades.Nowadays printed text recognition technology has becoming mature increasingly,and has been used widely in people's lives.However,improving the accuracy of text image of low quality in some certain application scenes has always been a bottleneck.Therefore,dealing with low accuracy which is caused by low quality characters is an issue to be solved.In the process of text image preprocessing,it's difficult to deal with merged characters and Chinese characters with left-right structure using traditional character projection segmentation method,this paper improves the character segmentation based on the projection method.Low quality text images often result in low resolution of character,to deal with this situation,image super-resolution technique based on the generative adversarial nets(GAN)is applied to fuzzy character images.By using a new perceptual loss function in the network instead of the traditional MSE loss function,character images can produce more rich texture details.Text recognition technique has evolved from the traditional feature extraction method to deep learning method.Convolutional neural network has solved many problems in computer vision field,and also provides a new end-to-end approach for text recognition.In this paper,a modified network based on GoogLeNet is used to deal with Chinese character set which is of a great variety and has a lots of similar characters in it.The network is simplified,but deep enough to do the recognition task.We use programs to generate massive training dataset,then train our network using Caffe framework,and finally get character classification model.Finally,in our experiments,we expand our training data by using various spatial transformation methods,and prove that data augmentation can improve recognition accuracy.Besides,we also prove that deep network has better performance for the classification task of Chinese characters.In addition,after the super-resolution processing of the characters,recognition accuracy has been improved to a certain degree.