ML-Based Prediction Enhancement Block for Video Compression

The patent introduces an intermediary ML-based prediction enhancement block that sits between the conventional predictive block and the residual calculation unit. This intermediary component uses machine learning models to generate enhanced prediction blocks from conventional prediction blocks, thereby improving compression efficiency without fundamentally redesigning the entire encoder architecture. The intermediary nature of this block allows it to leverage existing ML technologies while integrating them into the video coding pipeline.

The patent changes the prediction block generation process by introducing ML-based parameter transformation. Instead of using only conventional motion compensation parameters, the system transforms prediction blocks through ML models that learn optimal prediction parameters from training data. This parameter change approach allows the system to achieve better compression by adapting prediction parameters based on learned patterns from the video content.

The patent applies partial action by selectively applying machine learning-based prediction enhancement only to certain prediction blocks or specific coding units where it provides the most benefit. Rather than uniformly applying all available ML techniques to every prediction block, the system can apply enhancement selectively based on content characteristics, thereby improving prediction accuracy where needed while minimizing computational overhead in regions where conventional prediction is sufficient.

The system dynamically selects and switches between different machine learning techniques based on the characteristics of the input video content and prediction blocks. The prediction enhancement block can adapt its behavior by choosing different ML models or transformation methods depending on the current coding context, allowing optimal prediction accuracy to be achieved with minimal computational resources for each specific encoding scenario.