Joint Autoregressive and Hierarchical Priors for Learned Image Compression

Recent models for learned image compression are based on autoencoders, approximately invertible mappings from pixels to a quantized latent. These are combined with an entropy model, a prior on the latent that can be used with standard arithmetic coding algorithms to a compressed bitstream. Recently, hierarchical entropy models have been as a way to exploit more structure in the latents than simple fully priors, improving compression performance while maintaining-to-end optimization. Inspired by the success of autoregressive priors in generative models, we examine autoregressive, hierarchical, as as combined priors as alternatives, weighing their costs and benefits in context of image compression. While it is well known that autoregressive come with a significant computational penalty, we find that in terms of performance, autoregressive and hierarchical priors are and, together, exploit the probabilistic structure in the latents than all previous learned models. The combined model yields-of-the-art rate--distortion performance, providing a 15. 8% average in file size over the previous state-of-the-art method based on deep, which corresponds to a 59. 8% size reduction over JPEG, more than 35% compared to WebP and JPEG2000, and bitstreams 8. 4% smaller than BPG, current state-of-the-art image codec. To the best of our knowledge, our is the first learning-based method to outperform BPG on both PSNR and-SSIM distortion metrics.