Attention-Based Bi-Prediction Network for Versatile Video Coding (VVC) over 5G Network

Area of Science:

• Computer Science
• Electrical Engineering
• Information Technology

Background:

• Fifth-generation (5G) networks are crucial for emerging services like IoT, autonomous driving, and AR/VR.
• Versatile Video Coding (VVC) is a key standard for high-quality video compression.
• Existing bi-prediction methods in VVC, like BCW and BDOF, have limitations in representing diverse pixel variations and compensating bi-prediction blocks accurately.

Purpose of the Study:

• To propose an attention-based bi-prediction network (ABPN) as a superior alternative to current VVC bi-prediction methods.
• To improve the efficiency and accuracy of video compression for 5G applications.
• To develop a lightweight yet effective network for enhanced video coding.

Main Methods:

• Developed an attention-based bi-prediction network (ABPN) to learn efficient fused feature representations.
• Employed a knowledge distillation (KD) approach to create a compressed version of the ABPN.
• Integrated the lightweighted ABPN into the VTM-11.0 NNVC-1.0 standard reference software.

Main Results:

• The proposed ABPN effectively learns representations of fused features using an attention mechanism.
• The KD-based approach successfully compressed the ABPN size while maintaining comparable performance.
• Integration into VTM-11.0 resulted in significant BD-rate reductions: up to 5.89% (RA) and 4.91% (LDB) on the Y component compared to the VTM anchor.

Conclusions:

• The attention-based bi-prediction network (ABPN) offers a significant advancement over existing bi-prediction techniques in VVC.
• The lightweighted ABPN provides substantial coding gains, making it suitable for resource-constrained 5G environments.
• This approach effectively addresses the limitations of linear fusion and assumption-based methods in video compression.