Machine Learning-Based Fast Angular Prediction Mode Decision Technique in Video Coding.

Area of Science:

• Computer Vision
• Machine Learning
• Video Compression

Background:

• Intra-prediction mode decision is crucial for video compression efficiency.
• Current Rate-Distortion optimization is computationally intensive, increasing encoding time.
• Efficient mode decision is needed for real-time video coding applications.

Purpose of the Study:

• To develop a fast machine learning-based algorithm for intra-prediction mode decision.
• To reduce the computational complexity of video encoding processes.
• To integrate a novel approach into existing video coding standards like HEVC and JEM.

Main Methods:

• Utilized a random forest, an ensemble of randomized decision trees, for mode estimation.
• Developed a randomized tree model with parameterized split functions to learn block-based features.
• Employed a four-pixel feature set to capture directional block properties for fast evaluation.
• Integrated the inferred mode (IM) to reduce candidate modes before Rate-Distortion optimization.

Main Results:

• Achieved significant reduction in encoding time compared to reference software.
• Demonstrated only a slight increase in coding loss.
• Successfully integrated the random forest model into High Efficiency Video Coding Test Model (HM) and Joint Exploration Model (JEM).

Conclusions:

• The proposed machine learning approach offers a computationally efficient method for intra-prediction mode decision.
• This technique effectively balances encoding time reduction with minimal coding efficiency loss.
• The algorithm is suitable for integration into state-of-the-art video coding standards, enhancing performance.