Video Quality Model of Compression, Resolution and Frame Rate Adaptation Based on Space-Time Regularities

Area of Science:

• Video processing and quality assessment
• Computer vision
• Signal processing

Background:

• The streaming industry faces challenges predicting video quality due to increasing resolutions and frame rates.
• Dimension reduction techniques like spatial and temporal subsampling, combined with compression, significantly impact visual quality.
• Accurate prediction models are crucial for optimizing video delivery and user experience.

Purpose of the Study:

• To develop a novel video quality predictor sensitive to combined spatial, temporal, and compression distortions.
• To model space-time natural video statistics (NVS) for enhanced quality prediction.
• To achieve state-of-the-art (SOTA) prediction performance on challenging video quality databases.

Main Methods:

• Modeling statistics of divisively normalized difference between displaced neighboring frames.
• Extracting NVS features representing disturbances in space-time directional regularities.
• Training a regressor using parametric models of space-time regularities and extracted features.
• Testing the model on videos subjected to compression and space/time downsampling.

Main Results:

• Identified space-time paths with maximal regularity aligning with motion trajectories.
• Developed a video quality prediction engine based on disturbed space-time regularities.
• Achieved SOTA prediction performance on the ETRI-LIVE Space-Time Subsampled Video Quality (STSVQ) database.
• Demonstrated SOTA performance on the AVT-VQDB-UHD-1 database.

Conclusions:

• The proposed NVS-based predictor effectively models space-time distortions.
• The model accurately predicts perceptual video quality under combined compression and dimension reduction.
• This approach advances the field of video quality assessment for streaming applications.