Coupled Generative Models And Their Application To Image Pair Generation

With the development of deep learning,image generation techniques have made great progress,and researchers have proposed many image generation algorithms,but most image generation algorithms can only generate a single image,and some work requires the dataset support of image pairs.To address this problem,this paper focuses on the problem of generating image pairs.An image pair refers to a corresponding image tuple with the same high-level features of the subject and different low-level features,and a good image pair generated efficiently can better meet the needs of the relevant work tasks.There are various generative models,and the most studied ones are generative adversarial networks and variational autoencoders.This paper focuses on the coupled model of variational autoencoders and its variants.In this paper,two works are done for the generated image pair problem,the first is to improve the accuracy rate of the generated image pairs,and the second is to improve the resolution of the generated image pairs based on the first point.The main points are as follows:(1)A Coupled Variational Auto Encoder model Co VAE(coupled variational autoencoder)for improving the accuracy of generated image pairs is proposed.Existing methods require the presence of corresponding image pairs in different domains of the training set,but the coupled variational autoencoder does not require any corresponding image pairs to generate image pairs with different attributes.The coupled variational autoencoder is inspired by coupled generative adversarial networks and differs from the original method in that it learns high-level features with different attributes by training the coupled variational autoencoder to generate corresponding image pairs.The contributions of this work are three as follows: first,the coupled variational autoencoder is proposed to make an innovation in the model.Secondly,the image pairs are generated more accurately by learning high-level features of different attributes.Finally,the model is used to implement face attribute conversion as well as image interconversion.(2)A Coupled Adversarial Variational Auto Encoder model Co Ad VAE(coupled adversarial variational autoencoder)is proposed to improve the image resolution.Existing methods can generate accurate image pairs but not higher resolution pairs.Coupled adversarial variational autoencoders can generate higher resolution pairs more accurately.The previous method generates low resolution images due to the model limitations of the variational autoencoder.This paper adds an adversarial learning component to the original model and improves the loss function,which improves the resolution of the generated image pairs by introducing an adversarial learning mechanism based on Co VAE.The contributions of this work are as follows: First,the coupled adversarial variational autoencoder is proposed,which is an innovation in the model.Second,the loss function is improved to increase the resolution of the generated image pairs by adding an adversarial learning component.Thirdly,the model is extended to multiple datasets and its effectiveness in image attribute transformation is verified.Fourthly,the model is used to achieve a good image defogging effect.