Generative Adversarial Networks -based Reconstruction of Low Dimensional Autoencoder Representations

Abstract

Autoencoders comprise both an encoder and decoder, hence they hold the inherent capability of data compression and compressed sensing. However, the state-of-the-art for adversarial autoencoders (AAE) still cannot attain lossless recovery of original data, especially in the situation of low dimensional AAE representations. This paper proposes a general adversarial networks (GAN) based method to improve the reestablishment performance of AAEs. Adversarial learning is employed to generate high quality rebuilding sample which maximally approaches the original sample. The relationship between the dimensionality of AAE latent space and the reconstruction property is also explored. This paper focuses on the reconstruction of low dimensional AAE representation which is beneficial to signal, image and video coding and compressive sensing. It is demonstrated by the experimental results that the proposed method is superior to the AAE decoders in the recovery capability of original data.