MLISP: Esquema de Decisão ISP baseado em Aprendizado de Máquina para Codificadores VVC

Abstract

The Versatile Video Coding (VVC) standard introduces new coding tools to im prove compression efficiency, including the Intra Subpartition Prediction (ISP). Despite its benefits, ISP increases computational effort, as it adds new modes to be evaluated during the intra mode decision. To address this challenge, we propose MLISP, a machine learning-based scheme to accelerate this decision. MLISP consists of two complementary solutions. The first solution employs a decision tree trained with image features to predict whether the ISP evaluation is necessary for a given block. The second solution utilizes a decision tree trained with encoding features to determine which intra mode class should be considered if ISP evaluation is required. The experimental results demonstrate that the proposed approach reduces total encoding time by 10.97%, with a coding efficiency loss of only 0.32%. These results highlight the effectiveness of MLISP in accelerating encoding while maintaining a minimal impact on compression quality.