Multi-level ML Classifiers for Video Partition Search Early Termination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current video compression techniques, especially for high-definition clips, are computationally expensive and consume substantial resources due to exhaustive partition searches in video coding, which is inefficient for real-time applications over limited bandwidth channels.
Innovation Solution
Implementing a multi-level machine learning-based early termination scheme for partition searching in video coding, where classifiers are trained at different block size levels to determine whether to continue the search or terminate early, reducing computational complexity without sacrificing encoding quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If exhaustive partition search is used in video coding, then encoding quality is improved, but computational complexity and resource consumption increase significantly
Solution Approach 1:
The patent trains machine learning classifiers in advance on training video frames to learn optimal partitioning decisions. During encoding of target video frames, these pre-trained classifiers make rapid partitioning decisions without requiring exhaustive search, thus maintaining encoding quality while dramatically reducing computational complexity at runtime.
Solution Approach 2:
The patent uses training video frames as proxies to train classifiers that generalize to target video frames. The classifiers learn from the training data and can make accurate partitioning decisions on unseen target frames, avoiding the need to perform exhaustive searches on every frame while preserving encoding quality.
2Manufacturing precision
If exhaustive partition search is used in video coding, then encoding quality is improved, but processing speed decreases
Solution Approach 1:
The patent performs the computationally intensive work of learning optimal partitioning strategies in advance during the training phase. When encoding target video frames, the pre-trained classifiers make rapid decisions, achieving both high processing speed and maintained encoding quality without exhaustive searches.
Solution Approach 2:
The patent replaces the mechanical exhaustive search process with a machine learning-based decision system. The trained classifiers substitute for the traditional iterative search algorithm, providing rapid partitioning decisions that maintain quality while dramatically increasing processing speed for real-time applications.
3Productivity
If multi-level machine learning classifiers are implemented, then processing speed is improved, but device complexity increases
Solution Approach 1:
The patent divides the partitioning decision-making process into multiple levels, with each level handling different block sizes or partitioning scenarios. This segmentation allows the system to make rapid decisions at each level using specialized classifiers, achieving high processing speed while managing complexity through modular organization of the classification hierarchy.
Data Source
AI summary
Described herein are classifiers that are used to determine whether or not to partition a block in frame during prediction using recursive partitioning. Blocks of training video frames are encoded using recursive partitioning to generate encoded blocks. Training instances are generated for the encoded blocks that include values of features extracted from each encoded block and a label indicating whether or not the encoded block is partitioned into smaller blocks in the recursive partitioning. The classifiers are trained for different block sizes using the training instances associated with the block size as input to a machine-learning process. When encoding frames of a video sequence, the output of the classifiers determines whether input blocks are partitioned during encoding.