Multi-Dimensional Image Quality Assessment

• Multi-Dimensional Image Quality Assessment (MDIQA) is a framework that models quality as a vector of perceptual or task-specific dimensions, not a single score.
• It leverages explicit attribute prediction, multi-scale feature fusion, and multi-task deep learning to capture aspects like sharpness, color fidelity, and noise.
• MDIQA methods are applied in low-light enhancement, tone-mapped HDR, and medical imaging, improving interpretability and enabling adaptive restoration tuning.

Multi-Dimensional Image Quality Assessment (MDIQA) denotes a class of image quality assessment formulations in which quality is treated as a structured, multi-faceted phenomenon rather than as a single undifferentiated scalar. In the strongest sense, MDIQA predicts several perceptual dimensions explicitly—such as overall quality together with attribute-level scores—while broader usages also include frameworks that model multiple perceptual, representational, spatial, semantic, or task-specific dimensions internally before producing a final judgment (Yao et al., 23 Aug 2025, Sun et al., 29 Jun 2026, Chen et al., 7 Apr 2026).

1. Conceptual scope and boundary conditions

A central distinction in this literature is between true multidimensional output prediction and multi-feature or multi-method fusion yielding a single scalar overall quality score. This distinction is made explicitly in work on boosting-based IQA: predicting several perceptual axes separately is different from aggregating heterogeneous quality-sensitive estimators into one score. The latter may still be highly relevant to MDIQA, but it is not identical to it (Temel et al., 2018).

Within that distinction, at least three recurrent formulations appear. First, some methods produce explicit attribute vectors. LEIQ-Assessor predicts one overall MOS together with six perceptual sub-attributes—lightness, color fidelity, noise level, exposure quality, naturalness, and content recovery—using a multi-task setting designed for low-light enhanced images (Sun et al., 29 Jun 2026). MG-IQA extends reasoning-based NR-IQA from one scalar score to one overall score plus four attribute scores—sharpness, color fidelity, noise level, and composition—within a single inference pass (Chen et al., 7 Apr 2026). The unified MDIQA framework for evaluation and restoration models image quality using five technical and four aesthetic dimensions, with distinct branches and adaptive fusion into a final score (Yao et al., 23 Aug 2025).

Second, some methods are multidimensional in a structured latent sense. They operate over multiple scales, multiple feature depths, multiple spatial-angular axes, or multiple quality factors, but still regress to one final scalar. Tone-mapped HDR assessment based on multi-scale and multi-layer deep features is a representative example: it constructs a quality representation spanning local versus global information and low-, mid-, and high-level CNN features, then maps that representation to a single subjective score (He et al., 2018).

Third, some works broaden quality beyond human scalar opinion by redefining the target space itself. In machine-centric IQA, quality is explicitly decomposed into consistency and accuracy, and can further be studied across task, region, degradation type, and model ensemble. This is not HVS-centric MDIQA in the standard sense, but it is unmistakably multidimensional (Wang et al., 27 Aug 2025).

A concise way to view the field is therefore as a spectrum rather than a binary category.

Formulation	Representative works	Output form
Explicit attribute prediction	(Sun et al., 29 Jun 2026, Chen et al., 7 Apr 2026, Yao et al., 23 Aug 2025)	Overall score plus sub-attributes
Latent multidimensional representation	(He et al., 2018, Lan et al., 2024, Song et al., 2023)	Single scalar after structured fusion
Alternative multidimensional target spaces	(Wang et al., 27 Aug 2025, Liu et al., 2 Oct 2025)	Multiple machine or reasoning axes

This boundary matters because it prevents a common category error: not every multi-scale, multi-branch, or ensemble IQA model is a direct MDIQA model. Many are better understood as precursors, analogues, or partial realizations of the same underlying idea.

2. Historical lineage and methodological precursors

The broader IQA literature already implied that image quality is inherently multi-dimensional before MDIQA became explicit. A survey of subjective and objective IQA organized the field across multiple axes: subjective versus objective evaluation, full-/reduced-/no-reference settings, multiple distortion types, multiple datasets, and multiple image domains including conventional images, color images, HDR images, tone-mapped images, and 3-D images. It also showed that different metrics operationalize different perceptual emphases: SSIM separates luminance, contrast, and structure; MS-SSIM adds scale; VIF emphasizes information fidelity; MAD combines detection-based and appearance-based strategies; FSIM and FSIMc emphasize feature integrity and color; DRIM and TMQI address HDR- and tone-mapping-specific dimensions (Mohammadi et al., 2014). This survey did not define MDIQA as a formal framework, but it clearly established the conditions under which a one-number notion of quality becomes inadequate.

Deep scalar-IQA models then introduced architectural ingredients later absorbed by MDIQA. Multi-scale learning based on distorted images and residual maps combined spatial pyramid pooling and feature pyramids to capture local-to-global structure and hierarchical features, but remained a full-reference scalar regressor (Zhou et al., 2020). Full-reference transformer-based IQA combined multi-scale feature extraction with parallel transformer branches and simple cross-scale averaging, again strengthening representation without introducing explicit attribute outputs (Keshari et al., 2022). Multi-method boosting further showed that quality prediction improves when heterogeneous estimators—fidelity, perceptually extended fidelity, structural similarity, spectral similarity, color, and learning-based methods—are fused by a neural network or support vector machine, even though the output remains scalar (Temel et al., 2018).

A particularly important bridge to MDIQA is no-reference tone-mapped HDR IQA based on multi-scale and multi-layer deep features. That method explicitly argues that tone-mapped HDR quality depends on both local detail preservation and global contrast, and on both low-level appearance cues and higher-level content-related information. It uses original and downsampled images, extracts features from ResNet-50 layers res2a, res4b, and res4f, performs channel-wise mean and standard deviation pooling, concatenates them into a 4608-dimensional representation, and uses PLSR with 15 latent components for quality prediction (He et al., 2018). Although it still predicts one score, it makes the multidimensionality of perceptual evidence explicit.

These precursors collectively suggest that MDIQA did not emerge from a single architectural invention. It arose from the convergence of several lines of work: multi-scale processing, multi-branch feature extraction, multi-method fusion, domain-specific quality factorization, and the increasing availability of datasets with richer supervision.

3. Explicit multidimensional output models

The clearest direct formulation of MDIQA is the unified no-reference framework that models image quality with five technical and four aesthetic dimensions, trains a distinct branch for each dimension, and then fuses the resulting scores with an image-adaptive weighting branch and a 3-layer MLP to produce a final overall quality score (Yao et al., 23 Aug 2025). The architecture uses two shared backbones—one technical and one aesthetic—followed by dimension-specific heads, Gated Local Pooling from TOPIQ, Cross-Scale Attention Modules, and CLIP ResNet50 visual features injected in the second stage of training. Training is staged: the technical and aesthetic branches are first trained on the multi-dimensionally annotated SPAQ and PARA datasets, and the weighting branch and aggregation component are then optimized with frozen backbones and cross-scale attention modules. On authentic NR-IQA benchmarks, this model reaches the best reported results on KonIQ-10k and SPAQ, the best SRCC on FLIVE, and strong cross-dataset generalization. It is also repurposed as a tunable perceptual loss for Real-ESRGAN, where adjusting a dimension weight such as sharpness changes restoration behavior.

LEIQ-Assessor is a direct domain-specific MDIQA model for low-light enhanced images. It uses a pre-trained SigLIP2 Vision Transformer as a shared encoder and seven separate two-layer MLP regression heads to jointly predict overall MOS and six perceptual sub-attributes: lightness, color fidelity, noise level, exposure quality, naturalness, and content recovery (Sun et al., 29 Jun 2026). The quality space is explicitly defined as

$$\mathcal{T} = \{\text{mos},\, \text{light},\, \text{color},\, \text{noise},\, \text{exposure},\, \text{nature},\, \text{content\_recovery}\}.$$

The model is trained end-to-end on the MLE dataset of 800 low-light enhanced images with seven-dimensional labels, using Adam with learning rate $1\times 10^{-5}$, weight decay $1\times 10^{-7}$, batch size 8, and 30 epochs. The optimization target is an unweighted sum of per-dimension PLCC-based losses. On MLE, it substantially exceeds BRISQUE, NIQE, MANIQA, StairIQA, CLIP-IQA+, and LIQE on both overall MOS and each sub-attribute, and it achieved second place in the QoMEX 2026 Grand Challenge on Low-light Enhanced Image Quality Assessment.

MG-IQA pushes explicit multidimensionality into the reasoning regime. Built on Qwen2.5-VL-7B-Instruct, it extends reinforcement learning to rank from single-score IQA to joint prediction of overall quality and four fine-grained attributes—sharpness, color fidelity, noise level, and composition—within a single inference pass (Chen et al., 7 Apr 2026). Its attribute-aware prompt forces the model to reason step by step about each attribute and then synthesize them into an overall judgment. The main technical novelty is a multi-dimensional Thurstone reward model: $$P^{(a)}(x_i \succ x_j)= \Phi\left( \frac{\bar q^{(a)}(x_i)-\bar q^{(a)}(x_j)} {\sqrt{s^{2(a)}(x_i)+s^{2(a)}(x_j)}} \right),$$ combined with a composite reward

$$r_k(x_i) = w_0\, r_k^{\text{overall}}(x_i) + \sum_{a=1}^{A} w_a\, r_k^{(a)}(x_i),$$

and domain-adaptive weighting for synthetic, authentic, and AI-generated image datasets. Trained with attribute-aware SFT followed by GRPO-based RL2R, MG-IQA improves average SRCC by about $2.1\%$ over the previous reasoning-induced state of the art across eight IQA benchmarks, while also producing attribute-level scores and explicit rationales.

These three systems illustrate the main direct MDIQA pattern. Quality is represented as a vector, different components are supervised separately, and overall quality becomes an adaptive synthesis rather than the sole training target.

4. Structured multidimensionality without explicit attribute vectors

A substantial body of work stops short of explicit multi-output prediction but still models quality along several coupled dimensions. Such systems often function as methodological bridges between scalar IQA and full MDIQA.

For light field image quality assessment, Tensor-NLFQ treats the light field as a 4D tensor $\mathbf{L}(s,t,x,y)$ and separates quality into spatial-dimensional and angular-dimensional components (Zhou et al., 2019). Tucker decomposition yields first principal components for oriented sub-aperture stacks, from which the method derives the Principal Component Spatial Characteristic (PCSC) and the Tensor Angular Variation Index (TAVI). PCSC captures global naturalness and local frequency properties; TAVI captures angular consistency by modeling the structural-similarity profile between the first principal component and each view. Although the final output is a scalar quality score regressed by SVR, the method is directly multidimensional in its decomposition of the signal domain.

A related immersive-media example is full-reference stereoscopic 360-degree IQA from multi-viewports. That model decomposes quality into multiple perceptual components: viewport-space monocular image quality, binocular rivalry/dominance weighting between left and right views, depth-related information from the left-right difference map, and regression to both image quality and overall QoE (1908.10079). The binocular viewport quality is

$Q_n = w_n^L Q_n^L + w_n^R Q_n^R,$

and the depth proxy is the entropy of

$$\mathbf{S}_n^- = \left| \mathbf{I}_n^{DL} - \mathbf{I}_n^{DR} \right|.$$

This is multidimensional in representation and perceptual modeling, even though not all dimensions are output as separate final ratings.

AMqF is a full-reference method that explicitly decomposes deep features into quality factors such as luminance, contrast, and structure, then measures their responses in a learned dictionary space to better handle regional heterogeneity and non-uniform distortions (Lan et al., 2024). The approach remains a single-score predictor trained against MOS, but its internal representation is factorized, reconstruction-constrained, and decorrelated. In a strict taxonomy it is not a fully explicit MDIQA model; in practice it is a multi-factor perceptual IQA framework with strong MDIQA implications.

Medical LDCT IQA based on multi-scale distributions regression offers another implicit variant. MD-IQA predicts a scalar score, but the output is represented as several Gaussian-distributed target representations at different uncertainty scales, learned through a dual-branch ViT-plus-ConvNeXt network with deformable alignment, patch-wise weighted aggregation, and semi-supervised learning on unlabeled CT slices (Song et al., 2023). This is not attribute-level MDIQA, but it is explicitly multidimensional in output support, supervisory structure, and feature extraction.

AIGC image quality assessment has likewise moved toward multi-level and multi-task formulations. A multi-level visual representation paradigm for AIGC image quality separates perceptual quality and text-to-image correspondence into two task-specific networks: MGLF-Net, which fuses global Transformer and local CNN features for perceptual quality, and MPEF-Net, which injects prompt semantics into multi-level visual fusion for correspondence (Meng et al., 23 Jul 2025). On AGIQA-3K and AIGCIQA2023, this paradigm gives strong results on both tasks. It does not yet constitute a unified predictor over all AIGC quality axes, but it makes explicit that low-level perceptual defects and prompt-image alignment belong to different quality dimensions.

Across these cases, the field repeatedly returns to the same design principle: quality is easier to model when the representation preserves distinct axes—spatial versus angular, technical versus aesthetic, local versus global, or perceptual versus semantic—rather than collapsing them immediately.

5. Supervision, datasets, and evaluation protocols

MDIQA has expanded not only because of architectural changes but also because supervision has become richer. The supervision regimes now range from scalar MOS plus sub-attributes, to pairwise comparisons, to natural-language reasoning annotations, to machine-centric task-performance labels.

In low-light enhancement, the MLE dataset provides 800 low-light enhanced images with seven-dimensional quality labels, enabling direct multi-task regression over MOS and six perceptual attributes (Sun et al., 29 Jun 2026). In AIGC IQA, AGIQA-3K provides MOS for perceptual quality and text-to-image correspondence, while AIGCIQA2023 adds authenticity as a third dimension (Meng et al., 23 Jul 2025). In the unified MDIQA framework, multi-dimensional supervision comes from SPAQ and PARA, which provide the basis for technical and aesthetic branch pretraining (Yao et al., 23 Aug 2025).

An especially strong expansion of supervision appears in MedQ-Bench, which reformulates medical IQA as a perception–reasoning problem for multimodal LLMs (Liu et al., 2 Oct 2025). The benchmark spans five imaging modalities, over forty quality attributes, 2,600 perceptual queries, and 708 reasoning assessments. It defines two complementary tasks: MedQ-Perception, which probes low-level visual attribute perception through Yes-or-No, What, and How questions, and MedQ-Reasoning, which evaluates no-reference reasoning and pairwise comparison reasoning. Its judging protocol is explicitly multidimensional, scoring model outputs along completeness, preciseness, consistency, and quality accuracy rather than a single scalar. This benchmark demonstrates that MDIQA can refer not only to multiple output scores but also to multiple evaluative axes for generated reasoning.

Machine-centric IQA shifts the supervision target even more radically. MIQD-2.5M contains 2.5 million degraded samples generated from 10,000 source images, 10 degradation types, 5 severity levels, and 3 spatial region patterns, and labels each sample using aggregated downstream machine performance across 75 vision models and three tasks (Wang et al., 27 Aug 2025). The two primary quality dimensions are consistency and accuracy, with a composite score formed as

$$S(x', \mathcal{T}) = \lambda_1 \cdot \mathbb{E}_C(x, x') + \lambda_2 \cdot \mathbb{E}_A(x, x').$$

This makes the label space itself multidimensional and reveals that image quality can be task- and region-dependent in ways not captured by HVS-based MOS.

Evaluation protocols consequently vary. Many MDIQA-related systems still use SRCC and PLCC for scalar predictions or per-attribute regressions (Yao et al., 23 Aug 2025, Sun et al., 29 Jun 2026, Meng et al., 23 Jul 2025). Some add RMSE or KRCC depending on the benchmark (Song et al., 2023, Keshari et al., 2022). Reasoning-oriented systems require new metrics and human-alignment studies; MedQ-Bench validates automated multidimensional judging against specialists using quadratic weighted Cohen’s kappa (Liu et al., 2 Oct 2025). Machine-centric IQA requires separate reporting for consistency and accuracy, and also examines degradation-region sensitivity and cross-model stability (Wang et al., 27 Aug 2025).

This diversity of supervision suggests that MDIQA is not one benchmark problem but a family of related problems whose dimensions are partly determined by domain, task, and annotation protocol.

6. Applications, interpretability, and open issues

One reason MDIQA has attracted attention is that multidimensional prediction is useful beyond evaluation. The most explicit example is the unified MDIQA model for evaluation and restoration, which turns the trained assessor into a tunable perceptual loss for Real-ESRGAN (Yao et al., 23 Aug 2025). In no-reference mode, restoration maximizes predicted quality: $1\times 10^{-5}$0 while in full-reference mode it minimizes weighted dimension-specific feature distances: $1\times 10^{-5}$1 Because dimension weights are editable, restoration can be preference-aware. The paper explicitly shows that increasing the sharpness weight improves detail-oriented outputs, whereas increasing noisiness-related weight produces smoother outputs. It also shows that lower noise is not always equivalent to higher perceptual quality, an important corrective to overly simple scalar-fidelity assumptions.

Medical imaging has produced a parallel line of generalized, prompt-conditioned but still scalar assessors. MedIQA is presented as a comprehensive foundation model for medical IQA across 2D and 3D inputs, CT, MRI, and fundus data, multiple anatomical regions, and both physical-parameter pretraining and expert quality fine-tuning (Xun et al., 25 Jul 2025). Its automatic prompt strategy encodes dimensionality, modality, region, and type; its salient slice assessment module handles 3D volumes by selecting seven representative slices; and its MANIQA-based backbone aggregates patch- and slice-level predictions. This is not direct multi-output MDIQA, but it shows how multidimensional domain context can be incorporated even when the output remains scalar.

Another neighboring direction is unified FR/NR IQA with memory-driven distortion patterns. MQAF stores distortion prototypes in a memory bank and adaptively combines reference matching with memory-based distortion matching, allowing the same framework to operate in full-reference and no-reference modes (Lan et al., 20 Feb 2026). This is not attribute-level MDIQA, but it broadens the notion of multidimensionality to include multiple evidence sources and operating regimes.

Several open issues recur across the literature. A first is domain specificity. Many of the strongest direct MDIQA models are specialized to low-light enhancement, tone-mapped HDR, AIGC, medical imaging, or machine-centric assessment rather than presented as universal solutions (Sun et al., 29 Jun 2026, He et al., 2018, Meng et al., 23 Jul 2025, Liu et al., 2 Oct 2025, Wang et al., 27 Aug 2025). A second is label cost. Explicit attribute vectors, reasoning annotations, or machine-centric performance labels are all more expensive than overall MOS. A third is incomplete dimensional validation. Some systems use interpretable internal factors such as luminance, contrast, and structure without separately validating those factors against human sub-attribute judgments (Lan et al., 2024). A fourth is taxonomy instability: the relevant dimensions differ by application, and there is no single closed vocabulary spanning technical fidelity, aesthetics, semantics, diagnostic utility, and machine performance.

A persistent misconception is that MDIQA merely means “more complex scalar IQA.” The literature indicates a stricter and more useful view. In direct MDIQA, quality is modeled as a vector of perceptual or task-relevant dimensions, and overall quality is treated as an overview of those dimensions rather than the only object of prediction (Yao et al., 23 Aug 2025, Sun et al., 29 Jun 2026, Chen et al., 7 Apr 2026). In broader but still important variants, the same logic appears internally: separate quality factors, separate scales, separate view dimensions, separate semantic channels, or separate task-performance criteria are preserved long enough to influence the model’s reasoning. That shift—from scalar fitting to structured quality modeling—is the defining feature of the field.