A Research On Deep Learning Based Text Recognition And Generation In Natural Scene Images

Text is the main way for people to communicate.It has long been one of the technologies for researchers to enable computers to recognize text content in images and to generate images containing text content by computers.At the same time,it is of high practical value to accurately recognize and generate beautiful natural scene text images.However,the accuracy of the existing scene text recognition algorithms is still insufficient,and there is no mature technology for computer-generated scene text images.Therefore,by using deep learning technology,this paper proposes the recognition and generation method of natural scene characters based on the principle of deep learning.This paper proposes the following methods:First to address the problem of insufficient accuracy in text recognition of existing scenes,this paper improves on the existing Inception network and Dense Net network,and fuses feature maps extracted from the two networks.Convolution neural network can be used to extract data features effectively,but the network depth determines the ability of data feature extraction.According to the characteristics of convolutional neural networks,this paper proposes an improved fusion network structure that combines the existing Inception network and Dense Net networks.By designing different network structures and extracting different features of the image for fusion,the overall features and detailed features of the image can be effectively obtained.Secondly,it makes use of the characteristics of text sequences with context relations to extract text context information through Recurrent Neural Network(RNN);at the same time,it makes use of the Attention Mechanism to obtain effective text information;it accelerates the training process to achieve the purpose of improving recognition effect by improving Network structure.Second in order to solve the problem of automatic generation of scene text images by computer,this paper designs a Generative anti-neural network(Generative Adversarial Network GAN)to generate the scene text images.With the development of automation,making the computer automatically produce images containing artistic words can not only improve the efficiency of human generation but also solve the cost.Antagonistic neural network(GAN)is a technology that can automatically generate data.Conditional antagonistic neural network(CGAN)is an extension of GAN,which controls the generated results by adding a condition as auxiliary information,such as class tags or data from other modes.Therefore,on the basis of existing CGAN,this paper improves by adding word vectors as unified network training content toenable the computer to produce scene text images.At the same time,the existing gradient penalty Wasserstein GAN algorithm(WGAN-GP)is added in this paper,which can further improve the generation effect.At the same time,gated linear element(GLU)is adopted as the activation function in the network structure,which not only reduces the risk of gradient dispersion,but also effectively retains the nonlinear ability.Therefore,the overall ability of the computer to automatically generate scene text images is enhanced.