Neural Networks to estimate the probability of residual syntax elements for the av1 codec

Abstract

This work aims to showcase an Artificial Neural Network model to estimate the probability of residual syntax elements for the AV1 standard. Since the beginning of the 21st century, the usage of platforms that utilize video resources has been increasing. With the arrival of YouTube in 2005, it became possible to notice a growing usage of videos for entertainment and other mundane tasks. Over time, with the increase of social media and streaming applications, video traffic has become a large portion of global data traffic. In this context, it is recognized that there is a need for a video coding solution that can achieve a high level of compression. Codecs are a set of tools that are designed to significantly reduce the size of a video sequence. Although codecs that achieve a high compression rate are already available, many of them are locked down through royalty fees, which limits their usage. In 2018, the Alliance for Open Media (AOM) consortium created the codec AOMedia Video 1 (AV1) with the goal of providing a high-compression, royalty-free, open-source video format. As AV1 is a relatively new codec, finding new ways to improve its compression rate is a desired goal. In the AV1 context, residual syntax elements make up most of the entropy coded syntax elements, with entropy encoding being the last stage in the codec workflow. Therefore, accurately estimating their occurrence probabilities is crucial for obtaining a highly compressed bitstream (i.e., entropy encoding relies on these probabilities to compress the input data). To achieve this requisite, a total of 26 Artificial Neural Network models were evaluated, resulting in a final model that was capable of estimating the CDF of the DC_sign Syntax element, with an error of 0.11% when compared to the actual CDF.