Biprediction-Based Video Quality Enhancement via Learning.

Area of Science:

• Computer Vision
• Deep Learning
• Video Processing

Background:

• Traditional video quality enhancement methods using Convolutional Neural Networks (CNNs) often rely on optical flow for motion estimation, leading to high accuracy but also significant computational cost.
• The optical-flow-based method (OPT) is a common approach, but its complexity limits practical applications.

Purpose of the Study:

• To develop a novel, computationally efficient framework for multiframe video enhancement.
• To achieve high video quality comparable to existing methods while significantly reducing computational complexity.

Main Methods:

• Introduced a biprediction-based multiframe video enhancement (PMVE) framework.
• Designed a prediction network (Pred-net) to synthesize virtual frames (VFs) using frame pairs via biprediction.
• Developed a frame-fusion network (FF-net) to fuse VFs and low-quality frames (LFs), leveraging spatiotemporal correlations without explicit motion estimation.

Main Results:

• PMVE achieves peak signal-to-noise ratio (PSNR) performance on par with OPT.
• PMVE's computational complexity is only 1% of OPT.
• Outperforms other state-of-the-art methods in both objective and visual quality, with up to 0.42 dB PSNR improvement.

Conclusions:

• PMVE offers a highly effective and efficient solution for video quality enhancement.
• The proposed biprediction and frame-fusion strategy significantly reduces complexity while maintaining high enhancement accuracy.
• PMVE presents a practical advancement for real-time video enhancement applications.