Substitutional Neural Residual Compression for Video Encoding

The patent changes the fundamental parameters of video compression by replacing handcrafted linear tools with deep neural networks that can model nonlinear relationships. The DNNs learn optimal transformation parameters from data, enabling capture of nonlocal spatiotemporal correlations that traditional methods miss, thus improving compression performance despite increased model complexity.

The patent substitutes mechanical handcrafted coding tools with intelligent deep neural network systems. Instead of manually designed linear transforms and hand-tuned parameters, the system uses trained DNNs that automatically learn optimal compression representations, replacing the mechanical approach with an adaptive intelligent system.

The patent introduces dynamic adaptability through a single DNN model that can adjust its behavior based on target bitrate parameters. Instead of using multiple static model instances for different bitrates, the system uses one dynamic model that adapts its compression strategy based on the desired bitrate, reducing complexity while maintaining flexibility.

The patent creates a universal DNN model that can handle multiple bitrate requirements through a single instance. The model is designed to be multi-functional, capable of producing compressed representations at various bitrates by adjusting its internal parameters and transformation strategies, eliminating the need for separate specialized models for each bitrate target.