Neural-Network Video Filtering for Coding Efficiency and Performance

The patent applies preliminary action by performing filtering operations on reference video units before they are used for prediction of the current video unit. The filtering procedure is applied in advance to improve the quality of reference data, which subsequently enhances both coding efficiency and performance without requiring changes to the core coding architecture.

The patent utilizes parameter changes by dynamically adjusting filtering parameters based on the characteristics of the video content and the specific coding context. The filtering strength, kernel size, and other parameters are modified adaptively to optimize the balance between coding efficiency and performance for different video sequences and regions.

The patent applies local quality by selectively applying filtering procedures only to specific regions or video units where it provides the most benefit. Instead of uniformly filtering the entire video sequence, the method identifies and applies filtering locally to reference units that will have the greatest impact on prediction accuracy, thereby reducing overall computational complexity while maintaining performance improvements.

The patent employs partial action by applying filtering to only a subset of reference video units rather than all of them. The selection of which units to filter is based on their importance for prediction and their expected benefit from filtering, allowing the system to achieve performance improvement with reduced computational overhead compared to exhaustive filtering.