Video Processing With Neural Network Filters for Coding Efficiency

A neural network filter is introduced as an intermediary component between the video unit and the bitstream. The filter processes the video unit through multiple convolution layers and activation functions to generate enhanced color components, which are then used in the bitstream conversion process. This intermediary neural network filter improves both coding efficiency and effectiveness by optimizing the transformation between video units and bitstreams.

The patent applies parameter changes by transforming the video unit through neural network convolution operations that modify color component parameters. The neural network filter processes the video data through multiple layers with different kernel sizes and activation functions, changing the parameter representation of the video data to achieve better coding efficiency and effectiveness.

The neural network filter is segmented into multiple convolution layers, each performing a specific function. The filter includes an first convolution layer with first kernel size, a second convolution layer with second kernel size, and so on. This segmentation allows the complex neural network to be broken down into manageable modules that can be processed sequentially, reducing the overall system complexity while maintaining coding efficiency.

The patent applies partial action by selectively applying the neural network filter to specific color components (e.g., only to chroma components or only to luma components) rather than processing all components equally. This partial application reduces the computational burden and system complexity while still achieving significant coding efficiency improvements where most needed.