Combining Progressive Rethinking and Collaborative Learning: A Deep Framework for In-Loop Filtering.

Area of Science:

• Computer Vision
• Machine Learning
• Video Compression

Background:

• Deep learning models are increasingly used for video in-loop filtering.
• Existing methods face challenges in joint spatial-temporal modeling and side information injection.
• Optimizing reconstructed frame quality in video compression remains a key research area.

Purpose of the Study:

• To develop a deep learning-based in-loop filter addressing spatial-temporal modeling and side information injection.
• To enhance the quality of reconstructed intra-frames and inter-frames in video coding.
• To reduce the bitrate of compressed video sequences.

Main Methods:

• A Progressive Rethinking Network (PRN) for intra-frame coding, simulating human decision-making for spatial modeling.
• A collaborative learning mechanism for inter-frame coding, enabling feature-level interaction with reference frames.
• Extraction of intra-frame side information (partition map) and inter-frame side information (warped reference frame features).

Main Results:

• The PRN with intra-frame side information achieved a 9.0% average BD-rate reduction under All-intra (AI) configuration.
• The PRN with inter-frame side information achieved 9.0%, 10.6%, and 8.0% average BD-rate reductions under Low-Delay B (LDB), Low-Delay P (LDP), and Random Access (RA) configurations, respectively.
• Significant improvements in video compression efficiency were demonstrated compared to the HEVC baseline.

Conclusions:

• The proposed PRN effectively integrates spatial-temporal modeling and side information for improved video compression.
• The network's ability to progressively rethink and collaboratively learn enhances reconstructed frame quality.
• This deep learning approach offers a promising direction for next-generation video coding standards.