Bi-Directional Video Compression With Quantized Luma Motion Decoding

The patent replaces traditional mechanical video compression algorithms (motion estimation, loop filtering) with deep learning-based neural networks. The DVC framework uses end-to-end trainable neural networks to perform motion estimation and compensation, substituting complex iterative algorithms with learned models that achieve comparable or superior compression with reduced computational complexity during decoding.

The patent transforms the video compression problem by changing the fundamental parameters from hand-crafted algorithmic parameters to learnable neural network parameters. The system optimizes compression by training neural networks to learn optimal motion representation and compensation strategies, allowing adaptive parameter adjustment based on content characteristics rather than fixed algorithmic parameters.

The patent creates a universal DVC framework that integrates multiple compression functions (motion estimation, motion compensation, residual coding) into a single end-to-end trainable system. The neural network architecture performs multiple compression tasks simultaneously through shared layers and coordinated optimization, enabling global optimization across the entire compression pipeline rather than isolated module optimization.

The patent implements feedback mechanisms where the decoder's reconstruction quality information flows back to guide the encoder's compression decisions. The end-to-end training process uses reconstruction error feedback to adjust compression parameters and neural network weights, enabling iterative optimization that balances compression rate and quality across the entire system.

The patent transforms static pre-trained optical flow models into dynamic, adaptively trainable models. The DVC framework allows neural network parameters to be updated in real-time based on actual video content characteristics and compression requirements. The system dynamically adjusts motion estimation strategies by retraining or fine-tuning models during operation, enabling adaptation to different video sequences and compression scenarios.

The patent extends motion estimation from traditional uni-directional (past-to-future) to bi-directional (past-and-future-to-current) by incorporating future reference frames into the neural network's motion estimation process. This dimensional expansion in temporal reference usage enables more accurate motion prediction by leveraging information from both past and future frames simultaneously.