Tree-Structured Data Clustering-Driven Neural Network for Intra Prediction in Video Coding.

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• Intra prediction is vital for video compression, reducing spatial redundancy using local image information.
• Versatile Video Coding (H.266/VVC) utilizes directional prediction modes for intra prediction.
• Neural network-based methods have shown promise in enhancing video coding standards like HEVC and VVC.

Purpose of the Study:

• To propose a novel tree-structured data clustering-driven neural network (TreeNet) for intra prediction in video compression.
• To improve the efficiency and performance of intra prediction in the H.266/VVC standard.
• To investigate the effectiveness of TreeNet in assisting or replacing existing VVC intra prediction modes.

Main Methods:

• Developed TreeNet, a neural network that builds networks and clusters training data in a tree-structured manner.
• Each parent network splits into child networks, trained using hierarchically clustered data.
• Integrated TreeNet into VVC and proposed a fast termination strategy for accelerated search.

Main Results:

• When assisting VVC Intra modes, TreeNet (depth=3) achieved an average bitrate saving of 3.78% (up to 8.12%) over VTM-17.0.
• Replacing all VVC intra modes with TreeNet (depth=3) resulted in an average bitrate saving of 1.59%.
• TreeNet's hierarchical training enables distinct prediction and generalization abilities across network levels.

Conclusions:

• TreeNet offers a significant improvement for intra prediction in H.266/VVC.
• The proposed method effectively reduces bitrate, enhancing video compression efficiency.
• TreeNet presents a promising direction for future video coding standards.