Scalar Quality Metric (SQM) Overview

• Scalar Quality Metric (SQM) is a real-valued function that quantifies signal quality and reliability across various domains, including imaging, audio, and biomedical fields.
• SQMs employ techniques such as SVD energy retention, psychoacoustic filtering, and deep regression models to meet domain-specific quality assessments.
• They enable practical evaluation and optimization by establishing calibrated thresholds and integrating into loss functions for improved signal restoration and analysis.

A scalar quality metric (SQM) is a real-valued function that condenses the quality, reliability, or perceptual fidelity of a signal (image, audio, biomedical time series, or multidimensional data structures such as light fields) into a single, task-relevant number. SQMs are employed across multiple disciplines for objective assessment, evaluation, and algorithmic optimization, with mathematical definitions and methodologies tailored to the unique structural and perceptual properties of the target domain.

1. Mathematical Definitions and Domain-Specific Instantiations

Image Compression via SVD

In SVD-based image compression, the SQM $E_k$ quantifies the retained Hilbert–Schmidt norm (matrix Frobenius energy) in a rank-$k$ approximation:

$$E_k = \frac{\|A_k\|_F^2}{\|A\|_F^2} = \frac{\sum_{i=1}^k \sigma_i^2}{\sum_{i=1}^r \sigma_i^2}$$

where $A_k$ is the reconstructed image, $\sigma_i$ are singular values, and $r=\mathrm{rank}(A)$ (Razafindradina et al., 2017). $E_k$ thus measures preservation of global image information under SVD truncation and delineates perceptual quality zones (poor, good, very good) with thresholds at 99.9% and higher.

Sound and Speech Quality

In psychoacoustics, SQMs are used for attributes such as sharpness, roughness, and fluctuation strength, formulated as integrals or statistical summaries over time-frequency decompositions (e.g., via ISO 532-2 compliant gammatone/gammachirp filterbanks) (Isoyama et al., 2023), and in speech processing for direct prediction of mean-opinion-score (MOS) with deep architectures using large-scale pre-trained encoders (e.g., WhiSQA uses weighted Whisper-layer features and a Transformer regression head for MOS prediction) (Close et al., 4 Aug 2025).

Earth Observation Image Quality and No-Reference Metrics

QMRNet operationalizes SQMs in remote sensing by regressing blur, sharpness, SNR, relative edge response, and ground sampling distance from single image crops via a multi-head CNN architecture (Berga et al., 2022).

Sky Brightness and Environmental Monitoring

The Sky Quality Meter produces photometric SQMs as:

$$R_{\mathrm{SQM}} = \int L(\lambda) S_{\mathrm{SQM}}(\lambda) d\lambda$$

where $L(\lambda)$ is spectral radiance and $S_{\mathrm{SQM}}(\lambda)$ is spectral responsivity (Miguel et al., 2017). The output is transformed to surface brightness in AB magnitudes.

Task- and Metric-Specific Signal Quality Indices (SQI)

The pSQI formalizes signal quality for medical time series as the minimum downstream metric performance under bounded, task-driven perturbations:

$k$0

where $k$1 is the target algorithm, $k$2 is the performance metric, and $k$3 is an adversarial, energy-bounded perturbed version of $k$4 (Haidamous et al., 12 Feb 2026).

2. Methodological Taxonomy of Scalar Quality Metrics

SQMs are instantiated along several methodological axes:

Domain	Principle	Output Scalar
SVD Compression	Energy retention ratio	$k$5
Sound Quality	Psychoacoustic reduction	Sharpness, roughness, fluctuation strength
Speech Quality	Deep MOS regression	Normalized MOS
EO Image Quality	No-reference regression	Blur, SNR, sharpness, RER, GSD
Sky Brightness	Photometric integral	mag arcsec$k$6
Medical SQI	Task/metric perturbation	Worst-case $k$7

In all cases, the methodology involves (i) definition of quality with respect to digital or perceptual information retention or predictive task reliability, and (ii) rigorous calibration against either physical quantities, subjective human ratings, or algorithmic failure rates.

3. Comparative Analysis with Existing Metrics

Conventional metrics such as PSNR and SSIM, while efficient and widely adopted for image quality assessment, lack alignment with information preservation specific to the transform or compression scheme. The SVD-based $k$8 remains stable and tightly correlated with SVD compression perceptual quality, outperforming PSNR/SSIM in this context (Razafindradina et al., 2017). In EO applications and speech, deep learning-based SQMs tailored for perceptual or application-driven metrics generalize more robustly across distortion types than handcrafted or reference-dependent metrics (Berga et al., 2022, Close et al., 4 Aug 2025).

Whereas biomedical and environmental SQIs historically used feature-based or generic statistical attributes, pSQI's explicit alignment with algorithmic performance under perturbation allows for superior monotonicity and thresholded separation in practical downstream tasks (Haidamous et al., 12 Feb 2026).

4. Experimental Calibration and Quality Thresholds

Empirical calibration is domain-dependent but typically involves mapping SQM values to subjective or objective quality zones.

For SVD-image compression, quality regions are defined:

• Poor: $k$9 ≈ 0.994–0.9985, PSNR 27–34 dB, SSIM 0.82–0.93
• Good: $$E_k = \frac{\|A_k\|_F^2}{\|A\|_F^2} = \frac{\sum_{i=1}^k \sigma_i^2}{\sum_{i=1}^r \sigma_i^2}$$0 ≥ 0.9990, PSNR 35–42 dB, SSIM 0.94–0.98
• Very good: $$E_k = \frac{\|A_k\|_F^2}{\|A\|_F^2} = \frac{\sum_{i=1}^k \sigma_i^2}{\sum_{i=1}^r \sigma_i^2}$$1 ≥ 0.9999, PSNR ≥43 dB, SSIM 0.98–1.0

Speech SQMs (MOS) from WhiSQA demonstrate state-of-the-art correlation $$E_k = \frac{\|A_k\|_F^2}{\|A\|_F^2} = \frac{\sum_{i=1}^k \sigma_i^2}{\sum_{i=1}^r \sigma_i^2}$$2 (mean across test sets) and outperform prior MOS regression models (Close et al., 4 Aug 2025).

For medical time series, the pSQI achieves Spearman $$E_k = \frac{\|A_k\|_F^2}{\|A\|_F^2} = \frac{\sum_{i=1}^k \sigma_i^2}{\sum_{i=1}^r \sigma_i^2}$$3 and clear binary separation margins ($$E_k = \frac{\|A_k\|_F^2}{\|A\|_F^2} = \frac{\sum_{i=1}^k \sigma_i^2}{\sum_{i=1}^r \sigma_i^2}$$4 up to 0.37) in R-peak detection and atrial fibrillation classification, compared to feature-based or deep-learned alternatives (Haidamous et al., 12 Feb 2026).

5. Implementation Considerations and Limitations

• SVD-based SQMs require full or partial SVD, incurring $$E_k = \frac{\|A_k\|_F^2}{\|A\|_F^2} = \frac{\sum_{i=1}^k \sigma_i^2}{\sum_{i=1}^r \sigma_i^2}$$5 complexity; parallel or randomized SVD can mitigate this (Razafindradina et al., 2017).
• Psychoacoustic SQMs benefit from time-domain filterbanks (gammatone, gammachirp) for accurate, frame-level computation compliant with ISO 532-2 (Isoyama et al., 2023).
• Deep regression SQMs (QMRNet, WhiSQA) require large, labeled datasets and robust encoder architectures; data imbalance and domain adaptation are active concerns (Berga et al., 2022, Close et al., 4 Aug 2025).
• pSQI needs repeated invocations of the downstream algorithm under varied perturbations; optimization trade-offs exist between fidelity and computational tractability (Haidamous et al., 12 Feb 2026).
• Environmental measurement SQMs are confounded by spectral overlaps between device responsivity and changing sources (e.g., lamp spectra); color-dependent biases up to 1 magnitude can occur (Miguel et al., 2017).

6. Domain-Specific and Emerging Directions

Light field imaging exposes the need for 4D SQMs that jointly analyze spatial and angular error, unlike conventional 2D metrics which can achieve high correlation only in the presence of dense, undistorted references. Desired properties for new light-field SQMs include explicit modeling of spatio-angular gradients, angular coherence, perceptual linearity (e.g., just-objectionable-differences scale), and robustness to imperfect references (Adhikarla et al., 2017).

SQMs in modern pipelines may also serve as direct loss functions for model training (e.g., QMRLoss integrates scalar metrics into super-resolution objective functions), broadening their role from passive assessment to active optimization targets (Berga et al., 2022).

7. Practical Guidelines and Future Prospects

SQM deployment should match the metric's mathematical assumptions and domain calibration to the operational workflow. For SVD-based image compression, users should select $$E_k = \frac{\|A_k\|_F^2}{\|A\|_F^2} = \frac{\sum_{i=1}^k \sigma_i^2}{\sum_{i=1}^r \sigma_i^2}$$6 to achieve thresholded $$E_k = \frac{\|A_k\|_F^2}{\|A\|_F^2} = \frac{\sum_{i=1}^k \sigma_i^2}{\sum_{i=1}^r \sigma_i^2}$$7 values matching their acceptable quality zone (Razafindradina et al., 2017). In speech, leveraging foundation model features and attention architectures augments robustness and cross-domain generalization (Close et al., 4 Aug 2025). In biomedical signal analysis, pSQI tailors quality indices to the actual application and metric, enabling more precise input filtration and reliability control (Haidamous et al., 12 Feb 2026).

Anticipated developments include faster SVD and audio filterbank implementations, data-driven SQMs for high-dimensional and spatio-angular domains, gradient-based acceleration for task/metric SQIs, and further integration of SQMs as optimization criteria in restoration and enhancement networks across modalities.