Abstract
As video consumption grows across diverse devices and network conditions, ensuring a consistent Quality of Service (QoS) has become a critical challenge. Scalable Video Coding (SVC) offers a solution by providing a single bitstream that can be adapted spatially, temporally, and in quality. However, traditional objective metrics like PSNR often fail to reflect human perception. This paper proposes a universal quality metric for SVC that incorporates spatial and temporal video characteristics. By modeling the relationship between physical features and subjective viewer scores, the proposed metric achieves a 0.96 correlation with human perception, significantly outperforming standard methods.
The Challenge of Measuring Video Quality
In modern media delivery, video must be optimized for a vast range of displays, from mobile phones to high-resolution monitors. SVC is highly effective in these heterogeneous environments because it allows for flexible bitstream extraction.
Despite its technical advantages, a major hurdle remains: accurately measuring the "Quality of Experience" (QoE). Traditional metrics like Peak Signal-to-Noise Ratio (PSNR) are computationally simple but do not account for how humans actually perceive video. For instance, compression artifacts are often more noticeable in static scenes than in high-motion sequences. Existing metrics struggle to generalize across different content types, making it difficult for service providers to optimize delivery based on actual viewer satisfaction.
Proposed Methodology
The researchers developed a new Quality Metric (QM) designed to bridge the gap between objective data and subjective perception. The approach follows a structured pipeline:
- Feature Extraction: Using MPEG-7 descriptors, the system quantifies Spatial Value (SV) (texture complexity via Edge Histogram Descriptors) and Temporal Value (TV) (motion intensity via Motion Activity).
- Subjective Assessment: Experiments were conducted with 16 viewers to obtain Differential Mean Opinion Scores (DMOS) across various resolutions and bitrates, providing a human-centric ground truth.
- Modeling and Regression: Mathematical models were created to link SV, TV, frame rate, and resolution to subjective scores. The Least Square Method (LSM) was used to determine optimal coefficients for the final metric equation.
Content-Awareness and Subjective Perception
A core insight of this study is that subjective quality is not a fixed function of bitrate or PSNR; it is heavily influenced by the video's spatial characteristics.
As shown in Figure 1, for the same PSNR value, viewers perceive quality differently depending on the SV level. In high-complexity scenes (High SV), viewers are often more forgiving of certain artifacts, leading to higher subjective scores compared to low-complexity scenes at the same PSNR. This demonstrates why content-aware metrics are essential for accurate quality assessment.
Performance and Results
The proposed QM demonstrated superior performance and consistency compared to existing standards. While PSNR achieved a correlation of 0.79 and the Video Quality Metric (VQM) reached 0.90, the proposed QM achieved an average correlation coefficient of 0.96 with subjective human scores.
Figure 4 highlights the robustness of the proposed metric. While the default metric's accuracy fluctuates significantly depending on the video's spatial complexity, the proposed QM maintains a high and stable correlation (0.95–0.97) across all tested categories. This makes it a reliable tool for general-purpose video quality assessment.
Conclusion
By integrating spatial and temporal characteristics into the quality calculation, this research provides a robust framework for measuring perceived video quality in SVC environments. The high correlation with human judgment suggests that this metric can serve as an effective decision tool for SVC bitstream extraction, ensuring that users receive the best possible visual experience regardless of their device or network constraints. Future work may extend this model to account for network-induced issues like packet loss and jitter.
Summary updated