Introduction
Perceptual Visual Quality Assessment (PVQA) is the interdisciplinary field dedicated to developing computational models that can evaluate images and videos in a manner consistent with human perception. While traditional metrics like Peak Signal-to-Noise Ratio (PSNR) rely on pixel-level mathematical differences, they often fail to capture the subjective experience of a human viewer. As digital media evolves toward immersive virtual reality (VR) and generative AI (GenAI), the need for sophisticated, human-centric quality assessment has never been greater.
Framework of PVQA
The field of PVQA is broadly divided into two methodologies: subjective and objective. Subjective methods involve human observers rating content to establish a "gold standard," typically expressed as a Mean Opinion Score (MOS). Objective methods, conversely, are algorithms designed to predict these human scores automatically.
As shown in the figure above, objective methods are further categorized based on the availability of a reference signal:
- Full-Reference (FR): Compares a distorted image to a perfect original (e.g., SSIM, VMAF).
- Reduced-Reference (RR): Uses only partial features of the original.
- No-Reference (NR): Evaluates quality without any original "ground truth," which is essential for GenAI and live streaming applications.
The Role of the Human Visual System (HVS)
A core challenge in PVQA is modeling the biological nuances of the Human Visual System. Unlike a camera sensor, the HVS is highly adaptive. For instance, "masking" occurs when the eye ignores noise in high-texture areas, and sensitivity varies based on color and contrast. Modern PVQA models increasingly incorporate these HVS characteristics—such as sensitivity, color perception, and adaptation—to improve their predictive accuracy.
Emerging Frontiers: GenAI and Immersive Media
The rise of Generative AI has shifted the definition of "quality." In GenAI, assessment is not just about technical fidelity but also about semantic consistency (alignment with text prompts) and authenticity (avoiding the "uncanny valley"). Similarly, immersive media like 360-degree video and 3D Gaussian Splatting introduce complexities such as viewport-dependent quality and motion-induced sickness, which traditional 2D metrics are unequipped to handle.
Conclusion and Future Outlook
While Deep Learning and Foundation Models (like CLIP) have revolutionized the field, significant hurdles remain. Future research must address the high computational costs of Transformer-based models and the "black box" nature of neural networks. The ultimate goal is a universal foundation model capable of assessing any visual medium—from static images to complex 3D scenes—with human-like intuition.
Summary updated