RAM-VQA: Restoration Assisted Multi-Modality Video Quality Assessment.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Current video quality assessment (VQA) methods struggle with extreme quality videos and subtle variations.
• Vision-language models lack sensitivity to low-level distortions due to pre-training on high-level tasks.

Purpose of the Study:

• To develop a novel VQA framework, Restoration-Assisted Multi-modality VQA (RAM-VQA), that overcomes limitations of existing methods.
• To enhance VQA models' proficiency in evaluating videos at quality extremes and improve sensitivity to nuanced quality differences.

Main Methods:

• Proposed RAM-VQA framework leveraging video restoration to model distortion-sensitive features.
• Implemented a two-stage approach: prompt learning for quality-aware text space and dual-branch evaluation integrating semantic and technical cues.
• Utilized triple-level references (degraded, restored, pristine) and spatio-temporal differential analysis.

Main Results:

• RAM-VQA achieved state-of-the-art performance on diverse benchmarks.
• Demonstrated superior capability in handling severely degraded and near-perfect video content.
• Showcased robust generalization across different video quality scenarios.

Conclusions:

• The RAM-VQA framework effectively addresses limitations in current VQA methodologies.
• Video restoration as a proxy significantly improves the modeling of distortion-sensitive features for VQA.
• The proposed approach offers a more accurate and sensitive computational emulation of human perceptual judgments for video quality.