Learned Image Compression with Residual Coding

Abstract

We propose a two-layer image compression system consisting of a base-layer BPG codec and a learning-based residual layer codec. This proposal is submitted to the Challenge on Learned Image Compression (CLIC) in April 2019. Our contribution is to integrate several known components together to produce a result better than the original individual components. Also, unlike the conventional two-layer coding, our encoder and decoder take inputs also from the base-layer decoder. In addition, we create a refinement network to integrate the residual-layer decoded residual image and the base-layer decoded image together to form the final reconstructed image. Our simulation results indicate that the transmitted feature maps are fairly uncorrelated to the original image because the object boundary information can be provided by base-layer image. The experiments show that the proposed system achieves better performance than BPG subjectively at the given 0.15 bit-per-pixel constraint.