Full-Reference Objective Image Quality Assessment and Automated Artifact Detection: A Mathematical and Data-Driven Synthesis

The objective evaluation of image quality and the automated, spatially precise detection of visual artifacts remain among the most complex challenges in computational vision. Historically, the assessment of image fidelity relied on human visual inspection (Mean Opinion Score or MOS), which is subjective, costly, and fundamentally unscalable. The urgency for robust, 100% automated, data-driven computational models has accelerated exponentially with the advent of advanced generative architectures, such as text-to-image diffusion models and GANs. This paper introduces the Unified Probabilistic Image Quality and Artifact Locator (UPIQAL) framework, a novel, multi-tiered architecture designed to comprehensively replace human MOS. To satisfy the simultaneous requirements for deep semantic understanding, probabilistic spatial interpretability, precise color modeling, and deterministic artifact isolation, the UPIQAL algorithm synthesizes the dispersion-based texture tracking of A-DISTS, the multivariate uncertainty modeling of SUSS, the optimal transport color science of EDOKS, and targeted frequency-domain spatial heuristics. The UPIQAL algorithm processes reference and target images through five cascading mathematical modules: Module 1: Universal Preprocessing and Normalization: Ensures domain-invariant comparison using Minmax scaling and piece-wise linear histogram matching. Module 2: The Chromatic Transport Evaluator: Maps images to the perceptually uniform Oklab color space and utilizes the Earth Mover's Distance (EMD) via the Sinkhorn-Knopp algorithm to penalize pure color degradation. Module 3: The Hierarchical Deep Statistical Extractor: Extracts spatial statistics (mean, variance, cross-covariance) from a pre-trained VGG16 backbone, utilizing a spatial dispersion index to adaptively separate structural fidelity from textural similarity. Module 4: The Probabilistic Uncertainty Mapper: Models residual feature differences as samples from a multivariate Normal distribution, calculating the Mahalanobis distance to generate a highly precise Global Anomaly Map. Module 5: The Spatial Artifact Heuristics Engine: Deploys targeted mathematical heuristics — including a contrario cross-difference validation for JPEG blocking, local variance ratios for Gibbs ringing, and multi-level wavelet decomposition for Gaussian noise — to isolate specific degradations. Ultimately, the algorithm's aggregation head outputs an absolute objective scalar score calibrated to human psychophysics, alongside a multi-channel spatial diagnostic tensor that mathematically bounds and isolates generative and compression-based visual artifacts.