A Study of Deep-Learning-based Prediction Methods for Lossless Coding

In recent years, the research community started to explore new strategies for encoding image and video content based on innovative coding solutions developed using machine learning (ML) tools. An emerging research strategy proposes novel hybrid coding solutions as an alternative to traditional coding paradigms by replacing specific coding modules with efficient deep-learning (DL) based techniques. The paper presents a study on DL-based intra-prediction methods for lossless compression applications. For image coding, the paper studies our recently proposed pixel-wise prediction methods designed based on the residual learning concept, integrated into conventional lossless image coding frameworks. Moreover, a novel neural network design is proposed based on a new structure of layers. For video coding, the paper studies our recently proposed block-wise prediction methods designed based on recent breakthroughs in the ML domain, and integrated in the lossless HEVC standard. Experimental results show that the proposed lossless image codec achieves an improved performance with 1.6% compared to state-of-the-art DL-based methods. The study reveals that the hybrid coding solutions which incorporate DL-based prediction methods systematically and substantially improve the coding performance over traditional lossless coding paradigms.