Joint Machine Learning and Game Theory for Rate Control in High Efficiency Video Coding.

Area of Science:

• Video Compression
• Machine Learning
• Game Theory

Background:

• High Efficiency Video Coding (HEVC) faces challenges in optimizing bit allocation and rate control at the coding tree unit (CTU) level.
• Existing methods struggle with the rate-distortion (R-D) model prediction and the
• chicken-and-egg
• dilemma.

Purpose of the Study:

• To propose a novel joint machine learning and game theory (MLGT) framework for inter-frame CTU-level bit allocation and rate control (RC) optimization in HEVC.
• To improve prediction accuracy of the CTU-level R-D model and overcome the
• chicken-and-egg
• dilemma.

Main Methods:

• A support vector machine-based multi-classification scheme for enhanced R-D model prediction.
• A cooperative bargaining game theory approach for bit allocation optimization, proving utility function convexity and achieving Nash bargaining solution.
• Iterative solution search method with adjusted minimum utility based on reference coding distortion and quantization parameter (QP).

Main Results:

• The proposed MLGT-based RC method significantly improves R-D performance, quality smoothness, bit rate accuracy, and buffer control.
• Achieved R-D performance closely approaches the theoretical limits of the FixedQP method.
• Enhanced subjective visual quality compared to state-of-the-art one-pass RC methods.

Conclusions:

• The MLGT framework effectively optimizes inter-frame CTU-level bit allocation and rate control in HEVC.
• The approach successfully addresses R-D model prediction challenges and enhances overall video coding efficiency.
• The method offers superior performance in terms of quality, accuracy, and visual experience.