Research On Image Recognition Based On Generative Adversarial Network

With the rise of deep learning technology and the rapid development of hardware equipment,human beings are gradually entering the era of artificial intelligence.Image recognition is one of the most widely used technologies in the field of artificial intelligence.It is of great significance to study and improve the performance of image recognition.Based on the generative adversarial network,three network models are proposed in this paper,including SealGAN network,ESRGAN network and HWGAN network.Combined with automatic identification of invoice images,they are used for seal elimination,image super-resolution processing and enriching handwriting data sets,respectively.It can improve the performance of image recognition.The experimental results are compared and analyzed.The main research contents are as follows.Firstly,a new network architecture is propose,which is named SealGAN network and used for eliminating seals of invoice.Based on the cyclic structure of the CycleGAN network,the two independent classifiers are designed to replace the discriminative network in the SealGAN network.The generation network is designed based on the UNet and ResNet structures,which is adjusted based on the characteristics of the seal.Three types of network models,including the CycleGAN-Unet,CycleGAN-Resnet and SealGAN,have been used to perform invoice seal elimination.The network comprehensive performance evaluation indicators are proposed and used to evaluate these networks.The experimental results show the effectiveness of the SealGAN network.Secondly,typical image super-resolution processing methods are analyzed,including super-resolution based on traditional image processing,SRCNN model with deep convolutional neural network,and SRGAN model with generative adversarial network.Based on the conditional generative adversarial network cGAN,the auxiliary encoder is introduced and a new super-resolution model is proposed,which is named ESRGAN network.Combined with different invoice images,bilinear interpolation magnification,Sobel operator,Laplacian operator,SRCNN network,SRGAN network,and ESRGAN network are used to perform super-resolution processing experiments,and the processing effects of different methods are compared and analyzed.Finally,for the various rare characters appearing in handwritten Chinese characters,because of their small data set,it is impossible to train an automatic recognition network.The HWGAN network is designed,which is used to generate handwritten Chinese characters.It can enrich the handwritten Chinese character data set.The special training method is used in this network,and the PSNR index is introduced as a similarity loss to alleviate the model collapse defects of the original version of the generative adversarial network.The experimental results show that the HWGAN network can generate corresponding handwriting according to the printed body of Chinese characters,and the generated handwriting has different styles.