Inter-Coded Block Transform Using Directional LFNST

The non-separable transform is selectively applied only to specific inter-coded blocks that meet certain criteria (e.g., block size, prediction mode, gradient characteristics), rather than universally to all blocks. This segmentation approach captures directional information where beneficial while avoiding unnecessary computation in other cases, thus improving compression efficiency without proportionally increasing overall computational complexity.

The transform application is adapted locally based on block-specific characteristics such as gradient direction, block size, and prediction mode. By analyzing local features and applying the non-separable transform only where directional information is significant, the patent achieves improved compression for relevant blocks while maintaining lower complexity for blocks where simple transforms suffice.

Gradient calculations and directional analysis are performed preliminarily during the prediction stage, before the transform is applied. This preliminary assessment of block characteristics allows the system to quickly determine whether a non-separable transform is warranted, reducing the processing time overhead by avoiding unnecessary transform computations on blocks that don't benefit from directional modeling.

The patent applies the non-separable transform partially, only to the extent necessary for blocks with significant directional content. By using simple transforms or no transform for blocks without strong directional characteristics, the system achieves adequate coding accuracy for most blocks while reserving the computationally intensive non-separable transform only for cases where it provides meaningful improvement.