Research Of Image Feature Disentanglement And Multi-attribute Editing Based On Conditional Generative Adversarial Networks

Learning a generative model for the image is a critical problem in computer vision.The model aims to capture the image data distribution for generating new samples.However,since images in the high dimension space often lie in a complex manifold,it's a challenging task.With the development of deep learning,a series of deep generative models have emerged.Generative Adversarial Network(GAN)and Variational Auto-Encoder(VAE)are two of the most popular methods.To generate images based on a certain condition,the condition can be introduced into the model.This paper deals with conditional image generation task when the condition is a class label or an image,with the goal of enhancing the quality and diversity of generated images.The contents and contributions are as follows.1.When the condition is a class label,we propose to disentangle the latent space into label relevant and irrelevant dimensions by two separate encoders.Our model is based on the structure of CVAE-GAN,which is the hybrid model of Conditional Variational Auto-Encoder(CVAE)and Conditional Generative Adversarial Network(CGAN).To avoid the posterior collapse problem in VAE,Gaussian mixture distribution is introduced as the prior in place of the standard Gaussian.Each Gaussian component corresponds to a particular class,and the parameters of the Gaussian component are learned by label supervision.Experiments are conducted on the face recognition dataset and universal object dataset to show the improvements of our method on image quality and diversity.2.When the condition is an image,we employ two interactive decoders based on the structure of CGAN for face attribute editing.The first decoder provides mask and residual features for the second decoder at each intermediate layer to refine its features.In addition,a two stage training strategy is proposed to achieve fine control.In the first stage,the binary attribute labels are input as translation directions.In the second stage,the inputs indicate not only the directions,but also translation degrees.Also,domain consistency and source domain adversary loss are proposed for domain inter-polation training.Experiments on the face attribute dataset validate the effectiveness of our method.