Adaptive Chroma Prediction Based on Luma Difference for H.266/VVC.

Cross-component chroma prediction plays an important role in improving coding efficiency for H.266/VVC. We use the differences between reference samples and the predicted sample to design an attention model for chroma prediction, namely luma difference-based chroma prediction (LDCP). Specifically, the luma differences (LDs) between reference samples and the predicted sample are employed as the input of the attention model, which is designed as a softmax function to map LDs to chroma weights nonlinearly. Finally, a weighted chroma prediction is conducted based on the weights and chroma reference samples. To provide adaptive weights, the model parameter of the softmax function can be determined based on the template (T-LDCP) or offline learning (L-LDCP), respectively. Experimental results show that the T-LDCP achieves BD-rate reductions of 0.34%, 2.02%, and 2.34% for the Y, Cb, and Cr components, and the L-LDCP brings 0.32%, 2.06%, and 2.21% BD-rate savings for Y, Cb, and Cr components, respectively. The L-LDCP introduces slight encoding and decoding time increments, i.e., 2% and 1%, when integrated into the latest VVC test model version 18.0. Besides, the LDCP can be implemented by a pixel-level parallelization which is hardware-friendly.