ML Video In-Loop Filtering With Reduced Model Complexity

The patent extracts and removes partition information from the neural network input, keeping only the most essential features (reconstructed samples, predicted samples, boundary strength, quantization parameters). This extraction principle reduces the input data dimensionality and model complexity while preserving the critical information needed for effective filtering and compression.

The patent applies different filtering strategies to different regions of the video data by using boundary strength information to identify block boundaries. The neural network applies stronger filtering locally at boundaries where artifacts are more prominent, while using lighter filtering in smooth regions, thereby improving overall quality without uniformly increasing complexity across the entire video stream.

The patent explicitly removes partition information from the neural network input data, extracting only the essential features needed for filtering. This decision demonstrates the extraction principle by eliminating redundant information (partition details) that would increase processing complexity without providing sufficient additional benefit to filtering accuracy.

Instead of including partition information to improve filtering accuracy, the patent inverts the approach by excluding it and relying on alternative features like boundary strength and quantization parameters. This inversion shows that accurate filtering can be achieved without explicitly processing partition data, thereby reducing complexity while maintaining effectiveness.