Research On Image Translation Using A Generation Adversarial Network

Many cases have also proved that deep learning is a trend algorithm for the development of intellectual products.Deep learning is essentially data-driven.That is to say,the training model needs enough data to fit the accurate model.The quality of the data directly affects the quality of the model.This also determines that deep learning requires a large number of data sets for learning and modeling.Traditional data argmentation increases data diversity by flipping images,cropping images,and adding noise and so on.This method disadvantage is that the entire process is performed on existing training data.Image processing and computer vision are important research areas of the generative adversarial network(GAN).Data augmentation uses image translation,which can generate new images with labels or transform existing style data into other styles to expand the size of the training data.This thesis designs a new image translation framework based on the generative confrontation network,which can further generate effective high quality datasets.The main research contents are as follows:Based on GAN,a new ERGAN image translation framework is proposed.The framework of ERGAN image translation algorithm is constructed,Firstly,the model ideas of confrontation mechanism and dual learning(dual learning)are used to design duals GANs(28)?GA N_A,GAN_B?,so that the model can realize unsupervised learning image transaltion in unlabeled data sets.Secondly,in order to solve the common feature information between the unpaired input image and the generated sample,the reconstructed consistency loss function constraint is used,and the image reconstruction framework based on the reconstructed consistency loss function can find different domain feature relationships given the unpaired data.The experimental results show that the ERAGN framework can realize image translation and reconstruct the input images through generating image,which can learn the feature relations between different domains indirectly.Based on the research of generator network in the ERGAN image conversion framework,the image resolution of the traditional generator network architecture in GAN is relatively low,In order to make the generator network extract rich information features from the input image and retain the complete structure of the input image object.The generator network proposed in this thesis is composed of an encoder,a residual network(residual network),and a decoder.The experimental results show that using three unpaired data sets for training and testing models,compared our generator network architecture with the U-net and encoder-decode,the average PSNR/SSIM of the generated image by proposed generator network is increased by 7%/13%and 18%/43%,respectively,and it can generate 1024x1024high resolution image.Research on the stability of discriminator training based on ERGAN image transformation framework.The two discriminator networks in the ERGAN image transaltion framework are able to distinguish the similarity between the generated sample data and the input real image.The discriminator will supervise the generator to generate a more realistic image as the final image output.The existing image translation algorithm is face with discriminator instability during training,which causes the generator to no longer learn the input image feature information during the training process.The existing image translation algorithm is face with discriminator instability during training,which causes the generator to no longer learn the input image feature information during the training process.The result will seriously affect the quality of the generated sample image.Therefore,a new normalization algorithm:stable normalization(SN)is proposed as a layer of the discriminator network architecture in ERGAN to solve the discriminator instability problem.Finally,the different layers of the discriminator network are added to carry out the simulation experiment.The experimental results show that the image generated by the layer based on SN is more abundant images than other normalization methods,and the discriminator adds the SN algorithm to generate an image with a PSNR/SSIM value of 27.26/0.7414.Based on the above method,the ERGAN algorithm is trained and tested using six different data sets.The experimental results show that the image translation results of ERGAN are improved by 4.841dB/0.1934,0.81dB/0.0558,0.51dB/0.0638,1.18dB/0.0673,respectively,compared with Pix2pixGAN,CycleGAN,DualGAN,CoGAN image translation algorithms and so on.Average PSNR/SSIM of generating image on ERGAN algorithm is27.28/0.7414.ERGAN not only improves the quality of image translation results,but also implements multiple image transltion tasks.