Cross-Component Chroma Prediction Using Deep Learning Matrices

The patent replaces traditional mechanical video coding operations with deep learning-based neural network operations. Specifically, it uses a neural network model to predict chroma components from luma components, substituting conventional matrix-based chroma prediction methods with learning-based approaches that automatically learn optimal prediction patterns from training data, thereby improving coding efficiency while managing complexity through specialized hardware acceleration

The patent changes the fundamental parameters of video coding by introducing deep learning models with learnable weights and biases. Instead of using fixed prediction parameters, the system trains neural networks to optimize prediction performance dynamically, allowing the model parameters to adapt and improve coding efficiency across different video content types and scenarios

The patent introduces deep learning-based chroma prediction as an intermediary process between luma and chroma components. The neural network acts as a mediator that generates predicted chroma values from luma inputs, reducing the actual chroma data that needs to be stored and transmitted. This intermediary prediction mechanism allows high-resolution video to be compressed more efficiently while maintaining image quality

The patent creates a learned copy of chroma information from luma components through neural network prediction. Instead of storing and transmitting full chroma data at high resolution, the system generates a predictive copy of chroma values based on luma inputs and learned patterns, significantly reducing data volume while preserving visual quality through the redundancy exploitation capability of deep learning models