Research On Image Translation Method Based On Generative Adversarial Networ

In recent years,Generative Adversarial Network has become a research hotspot in the field of computer vision,and its training architecture based on adversarial thinking has performed well in many Image to Image Translation tasks.But there are also some limitations,such as uncontrollable attributes of the generated image,insufficient generalization of the generative model,and so on.This article focuses on the face to face and image super-resolution tasks in I2 I scenarios,and proposes two effective algorithm models.The specific work is as follows.First of all,this paper proposes a multi attribute face generation adaptive network based on potential coding space,which combines linear interpolation with GAN to achieve the task of face image translation and achieve good results.This method first designs a brand new encoder network to extract features from the input image and obtain feature vectors to form a potential encoding space;Secondly,based on the feature representation ability of the latent encoding space,the feature vectors are classified and regressed to obtain the attribute normal vector;Then linear interpolation method is used to combine the attribute normal vector with the original feature vector to obtain the newly sampled feature vector,and input it into the pre trained Style GAN to obtain the output image.In response to issues such as poor attribute control,loss of identity information,and poor model generalization during the face generation process,this method uses attribute normal vector retraining,decoupling,and other operations to process it,ultimately achieving multi attribute control of the face.The feasibility of this method was verified in the Sketch to Face task.This article uses the Celeb A-HQ dataset for training and testing,and selects facial image translation methods such as Interface GAN for comparative experiments and analysis.The results show that this method can generate high-resolution facial images while controlling their facial attributes.It can also effectively preserve the facial identity information of the input image,and has a certain degree of improvement compared to the original method in the evaluation indicators of FID,IS,and LPIPS,It is proved that the AF2F-GAN method is effective and progressiveness in the field of face image translation.Secondly,this paper proposes a Multi scale Few shot Super Resolution Generative Adversarial Network for SR tasks,which combines the residual idea of Residual Neural Network with GAN to achieve image super-resolution tasks and achieve good reconstruction results.This method constructs a brand new pyramid GAN network structure based on the residual idea,using multi-scale GAN.GANs of different scales correspond to images of different resolutions,and adopts a serial training structure.The output of the previous layer is trained as the input of the next layer after upsampling.This design can strengthen the model’s learning of the input image and enhance the detail generation effect of the image,Satisfactory results were achieved when training datasets for small samples.In this paper,classic superresolution methods such as SRGAN and ESRGAN are selected for comparative experiments and analysis,and evaluation indicators such as RMSE and PSNR are selected to prove the effectiveness of this method.The results show that the generation effect of this method is similar to that of SRGAN and other methods trained by a large number of data sets,and is significantly better than other small sample image super-resolution methods,which proves the effectiveness and progressiveness of MFSR-GAN in the field of small sample image super-resolution.