Temporal Referencing Network for Flicker-Free Video Processing

The patent implements a feedback mechanism by generating temporal reference maps from previous frame outputs and feeding them back as additional input to the image processing network for the current frame. This feedback loop allows the network to leverage temporal information and maintain consistency across frames, thereby reducing flicker while preserving colorization quality.

The patent performs preliminary action by pre-computing temporal reference maps from previous frames before processing the current frame. These reference maps are generated in advance and stored, allowing the image processing network to access temporal context without adding significant computational delay to the real-time processing pipeline.

The patent applies universality by designing the temporal referencing network to perform multiple functions: generating temporal reference maps, determining temporal relations, and converting image processing networks into video processing networks. This multi-functional approach consolidates what would otherwise require separate components into a single integrated system, managing complexity while achieving temporal stability.

The temporal referencing network acts as an intermediary layer between the image processing network and the final video output. It processes temporal information separately and then integrates it back into the image processing pipeline through the temporal reference maps, allowing the original image processing network to remain largely unchanged while gaining temporal stability capabilities.