TIMD Intra Prediction for Non-Dyadic Video Blocks

The patent segments the intra prediction process by distinguishing between dyadic and non-dyadic blocks. For non-dyadic blocks, it further segments the reference samples into multiple groups (first group from top boundary, second group from left boundary) to apply different processing strategies, thereby improving compression efficiency while managing complexity through structured segmentation.

The patent applies local quality by using different prediction approaches for different parts of the reference samples. Specifically, it applies gradient calculation and template matching to certain reference sample groups while using direct copying for others, optimizing the prediction quality locally for non-dyadic blocks without uniformly increasing complexity across all blocks.

The patent applies local quality by performing gradient calculation only on specific reference sample groups (first and second groups from top and left boundaries) for non-dyadic blocks, rather than uniformly across all reference samples. This localized approach improves prediction accuracy for the most critical regions while minimizing the overall computational load.

The patent segments the reference samples into multiple groups and applies gradient calculation selectively to these segments. By dividing the reference samples and applying computational operations only to specific segments rather than the entire set, the system achieves better prediction accuracy while managing computational resources more efficiently.