Machine Learning-Driven Multiple Transform Selection for Low-Complexity VVC Encoding

The H.266/VVC video coding standard achieves significant compression rates, but its deployment faces important challenges due to high computational demands. This is particularly evident in the Multiple Transform Selection (MTS) tool, which applies various combinations of transforms to enhance the energy concentration of the prediction residue. This work first presents an analysis of the MTS tool and shows that the distribution of transform mode decisions varies significantly. Next, a machine learning approach based on decision trees is employed to avoid the need for exhaustively testing all transform mode combinations in MTS. The predictive models are trained on features extracted directly from the encoding process, thus avoiding additional processing overhead. The proposed strategy reduces the overall encoding time by 30.32% in the All-Intra configuration, with only a minor increase of 1.17% in BD-rate, achieving a better balance between efficiency and complexity.