Content-Aware Convolutional Neural Network for In-Loop Filtering in High Efficiency Video Coding.

Area of Science:

• Computer Vision
• Machine Learning
• Video Compression

Background:

• Convolutional Neural Networks (CNNs) show great success in image processing.
• Improving video coding performance is crucial for efficient data transmission.

Purpose of the Study:

• To propose a content-aware CNN-based in-loop filter for High Efficiency Video Coding (HEVC).
• To enhance video coding performance through interpretable and optimal CNN model design.

Main Methods:

• A content-aware, multi-model filtering mechanism using CNNs applied to independent Coding Tree Units (CTUs).
• A discriminative neural network for adaptive CNN model selection based on region content characteristics.
• An iterative training method for CNN model parameter fine-tuning and filter category labeling.

Main Results:

• The proposed CNN-based in-loop filter significantly improves HEVC coding performance.
• Achieved up to 10.0% bit-rate reduction.
• Average bit-rate reductions of 4.1% to 6.0% across different HEVC configurations.

Conclusions:

• The content-aware CNN approach effectively enhances video coding efficiency in HEVC.
• The interpretable and adaptive CNN model design contributes to optimal in-loop filtering.
• This method offers a promising direction for next-generation video compression standards.