Adaptive Deep Reinforcement Learning-Based In-Loop Filter for VVC.

Area of Science:

• Computer Vision
• Machine Learning
• Video Compression

Background:

• Deep learning in-loop filters improve video coding efficiency and quality.
• Existing methods often use complex, non-adaptive single network structures.
• This limits performance across diverse video content.

Purpose of the Study:

• Propose an adaptive deep reinforcement learning-based in-loop filter (ARLF) for Versatile Video Coding (VVC).
• Enhance adaptability and performance of in-loop filtering in video coding.
• Achieve better rate-distortion optimization with reduced computational complexity.

Main Methods:

• Developed an adaptive deep reinforcement learning-based in-loop filter (ARLF).
• Treated filtering as a decision-making process using a deep reinforcement learning agent.
• Designed a lightweight backbone and a network set with varying complexities.
• Implemented a two-stage training scheme for robustness and an agent network for optimal selection.

Main Results:

• ARLF achieved average bitrate reductions of 2.17% (all-intra), 2.65% (low-delay P), 2.58% (low-delay), and 2.51% (random access).
• Demonstrated superior performance compared to existing deep learning-based methods.
• Showcased significantly lower computational complexity.

Conclusions:

• The proposed ARLF offers adaptive and efficient in-loop filtering for VVC.
• Deep reinforcement learning enables effective network selection for diverse video content.
• ARLF provides a promising direction for advanced video compression techniques.