Entropy Modeling in Video Compression Based on Machine Learning

Abstract

The paper investigates an entropy model of spatio-temporal characteristics in video compression using machine learning algorithms. A significant drawback of the vast majority of existing works on neural video codecs is the emphasis on the process of optimizing the creation of a latent representation through the study of various network structures, using well-studied models for image compression as an entropy model The model under study makes it possible to effectively evaluate both the spatial and temporal characteristics of compressed video data, which makes it possible to achieve a greater reduction in video redundancy. In addition, the universality of the entropy model also allows you to set the quantization step over the spatial channel. This content-adapted quantization mechanism, similar to rate control in standard codecs, not only helps achieve smooth compression ratio adjustment, but also improves final performance by dynamically distributing quantization intervals. The results of computational experiments on real video sequences confirm the efficiency of the studied video compression method.