A Learning-Based Framework for Low Bit-Rate Image and Video Coding

There is a major research effort under way to improve image and video coding efficiency through exploiting visual redundancy, in alignment with traditionally predictive coding and transform coding. It is motivated from the fact that natural images not only can be generally decomposed into texture and piecewise smooth parts called cartoon (e.g. edges), but may be recognized to consist of an overwhelming number of visual patterns generated by very diverse stochastic processes in nature. This paper explores perceptual non-parametric sampling methods into standardized video engine with structure-based prediction, and further suggests a learning-based framework for compressing image and video at low bit rate, by incorporating effective state-of-the-art inference algorithms to pursue an online synthesis solution. A crucial component is presented to learn the relationship (projection) between the abstracted patches (visual pattern) and the corresponding detail (feature space) in spatio-temporal manner. The experiment result shows the promising prospect for perceptual image and video coding.