Machine learning-based coding unit depth decisions for flexible complexity allocation in high efficiency video...

Area of Science:

• Computer Vision
• Machine Learning
• Video Compression

Background:

• High Efficiency Video Coding (HEVC) relies on complex coding unit (CU) depth decisions for rate-distortion (RD) optimization.
• Existing methods often face trade-offs between computational complexity and RD performance.

Purpose of the Study:

• To develop a machine learning-based method for fast CU depth decision in HEVC.
• To optimize complexity allocation under rate-distortion constraints.
• To improve the efficiency of HEVC video compression.

Main Methods:

• Modeling the quad-tree CU depth decision process as a hierarchical binary decision problem.
• Designing a flexible CU depth decision structure enabling smooth transfer between complexity and RD performance.
• Developing a three-output joint classifier to mitigate prediction risks.
• Deriving an RD-complexity model to optimize classifier parameters for minimal complexity under RD degradation constraints.

Main Results:

• Achieved computational complexity reduction ranging from 28.82% to 70.93% (51.45% on average) compared to the HEVC test model.
• Negligible impact on RD performance with an average Bjøntegaard-delta peak signal-to-noise ratio of -0.061 dB and Bjøntegaard-delta bit rate of 1.98%.
• Outperformed state-of-the-art schemes in overall performance.

Conclusions:

• The proposed machine learning-based CU depth decision method offers significant computational savings for HEVC.
• The algorithm effectively balances coding complexity and RD performance.
• This approach presents a promising advancement for efficient video compression.