Towards super resolution in the compressed domain of learning-based image codecs

Learning-based image coding has shown promising results during recent several years. Unlike the traditional approaches for image compression, learning-based codecs exploit deep neural networks for reducing dimensionality of the input at the stage where a linear transform would be typically applied previously. The signal representation after this stage, called latent space, caries the information in such a format that it can be interpreted by other deep neural networks without the need of decoding it. One of the tasks that can benefit from the above-mentioned possibility is super resolution. In this paper, we explore the possibilities and propose an approach for super resolution that is applied in the latent space. We focus on two types of architectures: fixed compression model and enhanced compression model. Additionally, we assess the performance of the proposed solutions.