Variational Autoencoder Based Image Compression with Pyramidal Features and Context Entropy Model

Abstract

Variational autoencoder with the potential to address an increasing need for flexible lossy image compression, has recently be investigated as a promising direction for advancing the state-of-the-art. Based on this effective framework, we present an end-to-end image compression method with a multi-scale encoder, residual decoder, and separate entropy model. The encoder uses a pyramidal resize module and inception network to leverage the priors at different resolution scales to improve the efficiency of the compressed latents. The decoder utilizes a residual network to synthesize the images with more nonlinearity. The separate entropy model is adopted to better predict the prior probability model of the latent representation. The final experiment results show that our approach yields a state-of-the-art image compression system.