Abstract
Over the past several years, we have witnessed the impressive progress of learned image compression. Recent learned image codecs are based on auto-encoders, that first encode an image into low-dimensional latent representations and then decode them for reconstruction. To capture spatial dependencies in the latent space, prior works exploit hyperprior and spatial context model to facilitate entropy estimation. However, they are hard to model effective long-range dependencies of the latents. In this paper, we explore to further reduce spatial redundancies among the latent variables by utilizing cross-channel relationships for explicit global prediction in the latent space. Obviously, it will generate bits overhead to transmit the prediction vectors that indicate the global correlations between reference point and current decoding point. Therefore, to avoid the transmission of overhead, we propose a 3-D global context model, which separates the latents into two channel groups. Once the first group is decoded, the proposed module will leverage the known group to model spatial correlations that guide the global prediction for the unknown group and thus achieve more efficient entropy estimation. Besides, we further adopt split attention module to build more powerful transform networks. Experimental results demonstrate that our full image compression model outperforms standard VVC/H.266 codec on Kodak dataset in terms of both PSNR and MS-SSIM, yielding the state-of-the-art rate-distortion performance.
Abstract (translated)
URL
https://arxiv.org/abs/2011.09704
