No-Reference Video Quality Assessment Using Multi-Pooled, Saliency Weighted Deep Features and Decision Fusion.

Area of Science:

• Computer Science
• Signal Processing
• Artificial Intelligence

Background:

• No-reference video quality assessment (NR-VQA) is crucial for evaluating perceptual quality in video services.
• Deep learning has emerged as a powerful tool for NR-VQA, driven by large benchmark databases.
• Existing methods face challenges in comprehensively characterizing diverse video distortions.

Purpose of the Study:

• To propose a novel deep learning-based approach for NR-VQA.
• To enhance the accuracy and robustness of perceptual quality assessment for videos.
• To leverage multiple pre-trained convolutional neural networks (CNNs) for versatile distortion characterization.

Main Methods:

• Utilizing a set of in-parallel pre-trained CNNs to extract deep features.
• Employing temporal pooling and saliency weighting for video-level feature representation.
• Mapping extracted features to perceptual quality scores independently using multiple regressors.
• Fusing quality scores from different regressors for final perceptual quality estimation.

Main Results:

• The proposed method achieves state-of-the-art performance on two large benchmark video quality assessment databases.
• Experimental results demonstrate the effectiveness of the approach on videos with authentic distortions.
• Decision fusion of multiple deep architectures significantly benefits NR-VQA performance.

Conclusions:

• The novel deep learning approach offers a significant advancement in NR-VQA.
• Fusion of multiple deep CNNs provides a robust and accurate method for perceptual quality assessment.
• The method sets a new benchmark for NR-VQA systems, particularly for authentic distortion scenarios.