CONVIQT: Contrastive Video Quality Estimator.

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Perceptual video quality assessment (VQA) is crucial for streaming platforms.
• Current methods often require labeled data for training.
• Developing self-supervised approaches for VQA is an active research area.

Purpose of the Study:

• To develop a self-supervised learning framework for perceptually relevant video quality representations.
• To train a deep learning model that captures both spatial and temporal video features.
• To evaluate the model's performance in a no-reference VQA setting.

Main Methods:

• A deep learning model combining a Convolutional Neural Network (CNN) for spatial features and a recurrent unit for temporal information was developed.
• Distortion type identification and degradation level determination were used as auxiliary tasks.
• The model, named CONtrastive VIdeo Quality EstimaTor (CONVIQT), was trained using a contrastive loss.
• Learned features were mapped to quality scores using a linear regressor in a no-reference setting.

Main Results:

• CONVIQT achieved competitive performance against state-of-the-art no-reference VQA models on multiple databases.
• The model demonstrated robustness and generalization across synthetic and realistic distortions.
• Ablation studies confirmed the effectiveness of the learned representations.

Conclusions:

• Self-supervised learning can yield compelling video quality representations with perceptual relevance.
• The CONVIQT framework offers a promising direction for efficient and effective VQA.
• The learned representations generalize well, reducing the need for extensive task-specific training data.