耦合变分自编码器及其在图像对生成中的应用

Coupled Variational Autoencoder and Its Application in Image Pair Generation

随着深度学习的发展,图像生成技术有了长足的进展,但大多数图像生成模型只能生成单一图像.针对这一问题,本文提出了一种耦合变分自编码器,它可以生成具有不同属性的人脸元组.现有的方法需要训练集的不同域中存在对应图像的元组,但是本文提出的方法不需要任何对应图像的元组,就可以生成具有不同属性的图像元组.本文的方法是在耦合生成对抗网络的灵感下提出的,与原有方法不同,它通过训练耦合变分自编码器模型来学习不同属性的特征表示,以生成对应图像元组.相比较原方法,它可以通过学习高级特征表示更精确的生成图像元组.此外,本文还用耦合变分自编码器实现了无监督人脸属性转换以及人脸的相互转换.将提出的方法应用于多个学习任务,包括生成不同属性的人脸元组、无监督的人脸属性转换以及图像相互转换.

With the development of deep learning,image generation technology has made great progress,but most image generation models can only generate a single image. In order to deal with this problem,a couple variational autoencoder is proposed,which can generate face tuples with different attributes. The existing method requires the existence of corresponding image tuples in different fields of the training set,but the method proposed in this paper can generate image tuples with different attributes without any corresponding image tuples. The method proposed in this paper is inspired by the coupled generative adversarial network,and different from the original method,it learns the feature representation of different attributes by training the coupled variational autoencoder model to generate corresponding image tuples. Compared with the original method,it can generate image tuples more accurately by learning high-level features. In addition,in this paper,the unsupervised face attribute conversion and face interconversion are also realized by couple variational autoencoder. The proposed method was applied to multiple learning tasks,including generating face tuples with different attributes,unsupervised face attribute conversion and image interconversion.