Neural Video Post-Processing for Resolution Downsampling

The patent extends the NNPF to perform multiple functions including both upsampling and downsampling operations. By configuring the NNPF with different purpose indicators (first indication for downsampling, second indication for upsampling), a single filter structure can adapt to various resolution conversion needs, eliminating the requirement for separate dedicated filters for each operation and thereby improving versatility without proportionally increasing system complexity

The patent introduces dynamic configuration mechanisms where the NNPF's operation mode (upsampling or downsampling) can be changed through purpose indications in the bitstream. This allows the system to adapt its processing capability dynamically based on the specific video content and requirements, enabling multi-functionality while maintaining a relatively simple base structure that doesn't require separate hardware for each operation type

The patent replaces traditional mechanical or algorithmic downsampling methods with a neural-network-based approach. The NNPF uses learned patterns and intelligent processing to perform downsampling while minimizing information loss, achieving better quality results compared to conventional methods. The neural network can be trained to preserve important visual information during resolution conversion, addressing the information loss problem while maintaining processing efficiency through optimized network inference