Neural Network Post Filter for Video Coding via SEI Indicators

The patent applies different filtering approaches to different video units based on their specific characteristics. Neural network filters are selectively applied only to video units that benefit from them, while other units use traditional filtering methods. This localized application improves video quality where needed without unnecessarily increasing bandwidth usage for all units.

The patent dynamically determines whether to apply neural network filters to each video unit based on real-time analysis of video characteristics, motion complexity, and other parameters. This dynamic decision-making process allows the system to adaptively optimize the balance between video quality improvement and bandwidth consumption, applying NN filters only when they provide meaningful benefits.

The patent applies neural network filters selectively to specific video units rather than uniformly across all units. By identifying which video units have characteristics that benefit from NN filtering (such as complex textures or difficult-to-compress regions), the system applies computational resources only where needed, reducing overall computational complexity while maintaining video quality improvements.

The patent applies neural network filtering partially, only to the extent necessary for achieving acceptable video quality. Rather than applying NN filters to all video units regardless of need, the system performs partial application based on video content analysis, thereby reducing unnecessary computational complexity while still achieving quality improvements where they matter most.

The patent segments the filter model indication information into different levels of detail. Rather than including complete filter model specifications for every video unit, the system uses a hierarchical approach where general filter model selections are indicated at higher levels (e.g., sequence or picture level) and specific adjustments are made only where necessary. This segmentation reduces bitstream overhead while maintaining adaptability for selective video units.

The patent applies filter model indication partially, providing detailed model selection information only for video units that require neural network filtering. For units that use standard filtering or default NN models, the system omits or simplifies the indication data. This partial information provision maintains adaptability for units that need it while significantly reducing overall bitstream overhead.