In-Loop Sub-Image-Level Noise Modeling for Realistic Video Compression

The patent divides the image frame into multiple regions (e.g., foreground, background, different depth zones) and applies different noise models to each region. This segmentation allows the system to maintain compression efficiency through structured processing while preserving realistic noise characteristics in each specific region, resolving the contradiction between compression efficiency and noise accuracy.

The patent implements location-dependent noise modeling where different noise parameters (variance, correlation, temporal characteristics) are applied to different spatial locations within the frame. This local quality approach ensures that noise characteristics match the specific content of each region, maintaining both compression efficiency through systematic processing and realistic noise reproduction through localized adaptation.

The patent performs noise parameter estimation and region classification in advance during the encoding process, storing these parameters for use during decoding. By preparing noise models beforehand based on content analysis, the system achieves realistic noise generation without adding significant complexity to the main compression loop, as the heavy lifting is done during the initial pass.

The patent implements self-adaptive noise modeling where the system automatically estimates noise parameters from the video content itself and generates appropriate noise models without requiring external input or manual configuration. This self-service capability reduces operational complexity while maintaining high noise realism through content-driven adaptation.