Hierarchical Image Compression Framework

Yunying Ge, Jing Wang, Yibo Shi, Shangyin Gao

Keywords: Image compression, Hierarchical Compression, Hierarchical Condition, Residual Compression

TL;DR: A Hierarchical Image Compression framework can break the limitation of VAE and achieve much better performance at high bit rates with smaller FLOPs.

Abstract: In learning-based image compression approaches, compression models are based on variational autoencoder(VAE) framework and optimized by a rate-distortion objective function, which achieve better performance than hybrid codecs. However, VAE maps the input to a lower dimensional latent space which becomes a bottleneck of reconstruction. In this paper, we propose a deep Hierarchical Compression(HC) model, which can achieve good compression performance from low-bit to very high-bit. HC model consists of two closely-related modules, including hierarchical latent compression module and Hierarchical Conditional Entropy(HCE) module. Such a design transmits the details in the shallower layers and coarse information in the deeper layers and conditions the shallower entropy estimation on the deeper information. Extensive experiments show that HC model could breakthrough the AE limit and achieve significant improvements over state-of-the-art approaches in the high quality regime.