Refined Chroma From Luma Prediction in AV1 Based on Color Component Grouping

Junyan Huo, Zhenyao Zhang, Jiarun Song, Yanzhuo Ma, Fuzheng Yang

Abstract: Chroma from luma (CfL) in AOMedia Video 1 (AV1) utilizes the correlation between color components to derive the predicted chroma from the reconstructed luma. The mean of the predicted chroma, i.e., dc, is set to the chroma average of the neighboring regions. Since the neighboring regions have limited reference samples, the distribution of the coding block is hard to be predicted from the neighboring regions, resulting in an error in the predicted dc. A new feature of CfL is that the luma of the coding block has been reconstructed. Using the luma information, the ground-truth distribution of the coding block can be built. Based on this observation, a refined dc prediction is proposed based on color component grouping (GDC). We design an adaptive grouping scheme and use the number of luma and the chroma average in each group to derive a refined dc. The offline experiment verifies that the proposed method provides a dc with high prediction accuracy. Compared with the libaom of AV1, the proposed GDC achieves BD-rate reductions of 0.15%, 0.90%, and 0.98% for the Y, Cb, and Cr components. With the increased group numbers, additional coding gains can be provided. The proposed GDC can be implemented with a small complexity, which is hardware-friendly.