Face Image Quality Assessment (FIQA)

• Face Image Quality Assessment is the estimation of a face image's biometric utility by evaluating factors like pose, illumination, and blur.
• The field employs both factor-specific and monolithic approaches, including deep learning, perturbation analysis, and transformer-based mechanisms.
• FIQA is applied to filter low-quality images, reduce matching errors, guide reacquisition, and enhance overall recognition performance.

Face Image Quality Assessment (FIQA) estimates the utility of a captured face image for automated face recognition (FR) systems. In the literature, “quality” is commonly differentiated into character, fidelity, and utility, with FIQA primarily targeting biometric utility: the suitability of an image for reliable recognition rather than merely its visual appeal. In operational terms, FIQA is used to reject low-quality samples, reduce false match and false non-match errors, guide reacquisition, support template selection, and stabilize recognition in unconstrained settings. The field has evolved from factor-specific measurements of pose, illumination, blur, and resolution to monolithic deep predictors, FR-integrated quality models, training-free probes of model robustness, and transformer-based mechanisms that operate on attention or patch dynamics (Schlett et al., 2020, Boutros et al., 2021, Babnik et al., 2023).

1. Conceptual foundations of face image quality

The survey literature treats biometric utility as the central FIQA notion. Under this view, a face image is high quality if it is expected to support accurate matching by a target FR system; it is low quality if it increases recognition errors or causes unstable similarity scores. This formulation explains why FIQA is tightly coupled to FR backbones, training data, and operating points rather than to generic perceptual quality alone (Schlett et al., 2020).

The principal factors reported to affect face image utility are pose, illumination, blur, resolution, expression, occlusion, image noise, skin color, background, and compression artifacts. These factors degrade either the fidelity of the face signal or the discriminability of the resulting embedding. In constrained settings, modern FR systems may approach near-ideal performance, but FIQA becomes especially important in unconstrained scenarios, where the captured data exhibit large variations in quality and acquisition conditions (Schlett et al., 2020, Babnik et al., 2023).

Beyond image filtering, FIQA has been described as useful for immediate feedback at acquisition, quality summarization/monitoring, progressive identification, quality-weighted score fusion, database maintenance, adaptive enhancement or compression, context switching between matchers, and partial presentation attack defense. This breadth of use has contributed to the coexistence of multiple FIQA definitions and evaluation protocols, including both biometric-utility-oriented and perceptual-quality-oriented formulations (Schlett et al., 2020).

2. Taxonomy and major methodological paradigms

A widely used taxonomy divides FIQA into factor-specific approaches and monolithic approaches. Factor-specific methods target explicit, interpretable quality factors such as blur or illumination, while monolithic methods output a single quality score, often through a deep model. The survey further distinguishes handcrafted, utility-agnostic training, human-ground-truth-trained, FR-based ground-truth-trained, FR-based inference, and FR-integration paradigms, as well as explicit or trained fusion methods that combine multiple factor scores (Schlett et al., 2020).

A complementary grouping separates general-purpose image quality assessment (IQA) from face-specific FIQA. In a bias study, the evaluated families were: General-purpose IQA with BRISQUE, NIQE, and RankIQA; Supervised Face Image Quality Assessment (sFIQA) with SDD-FIQA, MagFace, and CR-FIQA; and Unsupervised Face Image Quality Assessment (uFIQA) with SER-FIQ and FaceQAN (Babnik et al., 2022).

Paradigm	Representative methods	Core signal
Factor-specific / handcrafted	histogram analysis for illumination; edge detection for blur; inter-eye distance as size proxy	explicit quality factors
FR-integrated	CR-FIQA, MagFace, PFE	classifiability, magnitude, or uncertainty in embedding space
Distributional pseudo-labeling	SDD-FIQA, IG-FIQA	intra-class and inter-class similarity structure
Training-free robustness probes	DifFIQA, GraFIQs, PreFIQs, FROQ	perturbation stability, BN-statistic adaptation, pruning drift, intermediate activations
Transformer-based	ViT-FIQA, ATTN-FIQA, ViTNT-FIQA	quality token, attention magnitudes, patch-evolution stability
Generic/perceptual face IQA	DSL-FIQA, VQualA methods	MOS prediction and degradation modeling

This taxonomy makes clear that FIQA is not a single technique but a family of estimation strategies tied to different assumptions about what quality should encode. A plausible implication is that methodological differences often reflect different target definitions: biometric utility for FR on the one hand, and perceptual facial quality or Mean Opinion Scores (MOS) on the other (Schlett et al., 2020, Chen et al., 2024, Ma et al., 25 Aug 2025).

3. Classifiability, similarity structure, and pseudo-label learning

A major line of work learns FIQA from geometric or statistical structure already present in FR training. CR-FIQA defines face image quality through sample relative classifiability. In normalized feature space, it uses the Class Center Angular Similarity,

$$CCS_{x_i} = \cos(\theta_{y_i}) = x_i \cdot w_{y_i},$$

the Nearest Negative Class Center Angular Similarity,

$NNCCS_{x_i} = \max_{j \ne y_i} x_i \cdot w_j,$

and the Certainty Ratio,

$$CR_{x_i} = \frac{CCS_{x_i}}{NNCCS_{x_i} + (1 + \epsilon)}.$$

A regression branch is trained jointly with ArcFace loss so that the network predicts this relative classifiability on unseen identities; empirically, CR-FIQA achieved the best performance in almost all protocol/model/benchmark combinations reported in its study (Boutros et al., 2021).

SDD-FIQA starts from the statement that “a high-quality face image should be similar to its intra-class samples and dissimilar to its inter-class samples.” It builds per-sample positive and negative similarity sets and uses the Wasserstein Distance (WD) between the intra-class and inter-class similarity distributions as an unsupervised pseudo-label: $$Q_{x_i} = \mathbb{WD}\left(\mathcal{S}^P_{x_i} \,\|\, \mathcal{S}^N_{x_i}\right).$$ The labels are normalized to $[0,100]$, a regression network is trained with Huber loss, and a fast approximation samples positive and negative similarities repeatedly to reduce complexity from $\mathcal{O}(n^2)$ to $\mathcal{O}(n)$. The reported experiments showed strong gains over prior baselines and good cross-model generalization (Ou et al., 2021).

Later work targeted the quality of these pseudo-labels themselves. IG-FIQA observes that pseudo-labels from classes with low intra-class variance may be unrelated to actual quality, because even poor-quality samples from highly homogeneous classes may receive high scores. It estimates class-wise variance online via an EMA of $1-CCS_{x_i}$,

$$v_{y_i}^t = \alpha \cdot v_{y_i}^{t-1} + (1-\alpha)\cdot(1-CCS_{x_i}),$$

then uses a z-score-based weighting mask so that the lowest-variance classes can contribute little or nothing to the regression loss. Its training also separates clean images for the FR backbone from augmented images for the regression branch, with rescaling, random erasing, color jittering, and horizontal flip applied to the latter; the method reported novel state-of-the-art performance on several benchmarks (Kim et al., 2024).

A related refinement strategy is the supervised label-optimization procedure that takes an existing FIQA method $Q$, combines its scores with mated pair similarities from a selected FR model, and reorders the quality distribution by optimizing the score indices. The method preserves the original score distribution while improving fine-grained ranking, and it was evaluated as an add-on to CR-FIQA, FaceQAN, SER-FIQ, PCNet, MagFace, and SDD-FIQA (Babnik et al., 2023).

4. Training-free FIQA through perturbation, adaptation, and model sensitivity

Another dominant direction dispenses with explicit FIQA training and instead probes how a pre-trained FR model reacts to controlled perturbations. DifFIQA uses denoising diffusion probabilistic models (DDPMs) and measures embedding stability under both the forward noising process and the backward denoising process. Original and horizontally flipped face images are perturbed, embedded, and scored by the average cosine similarity between the original embedding and the perturbed embeddings across multiple random diffusion runs. DifFIQA and its distilled regressor DifFIQA(R) were evaluated on 7 datasets, with 4 target FR models, against 10 state-of-the-art FIQA techniques, and were consistently among the top methods; on CosFace at drop rate $NNCCS_{x_i} = \max_{j \ne y_i} x_i \cdot w_j,$0, the reported average pAUC values were 0.675 for DifFIQA, 0.669 for DifFIQA(R), and 0.678 for the best prior SOTA CR-FIQA (Babnik et al., 2023).

GraFIQs measures the discrepancy between the Batch Normalization statistics (BNS) stored during FR training and those induced by a test image. For each BN layer, it computes the MSE between stored and observed mean/variance, backpropagates this loss through the frozen FR model, and uses the cumulative absolute sum of the resulting gradient magnitudes as the FIQA signal: $NNCCS_{x_i} = \max_{j \ne y_i} x_i \cdot w_j,$1 The method is training-free and quality-labeling-free, requires one forward and one backward pass, and was reported to outperform general IQA methods while remaining competitive with recent FIQA approaches, especially on XQLFW (Kolf et al., 2024).

PreFIQs formulates the Pruning Identified Exemplar (PIE) hypothesis: low-utility images rely disproportionately on fragile network parameters and therefore show larger embedding drift under pruning. Using a pretrained FR model $NNCCS_{x_i} = \max_{j \ne y_i} x_i \cdot w_j,$2 and its pruned version $NNCCS_{x_i} = \max_{j \ne y_i} x_i \cdot w_j,$3, it defines drift

$NNCCS_{x_i} = \max_{j \ne y_i} x_i \cdot w_j,$4

and maps it to quality by

$NNCCS_{x_i} = \max_{j \ne y_i} x_i \cdot w_j,$5

The paper provides a first-order justification through a Jacobian-vector product analysis and reports competitive or superior performance across eight benchmarks and four FR models without any training or supervision (Kolf et al., 13 May 2026).

FROQ takes a different training-free route. It treats intermediate activations of a given FR model as a Quality Observer, defines an aggregation $NNCCS_{x_i} = \max_{j \ne y_i} x_i \cdot w_j,$6 as the $NNCCS_{x_i} = \max_{j \ne y_i} x_i \cdot w_j,$7 norm of flattened activations, and averages the selected layer responses,

$NNCCS_{x_i} = \max_{j \ne y_i} x_i \cdot w_j,$8

A calibration step on a small set of images uses pseudo-labels from a perturbation-based unsupervised FIQA method—horizontal flip, Gaussian noise, and partial occlusion scored by cosine similarity—to identify the most informative layers by Spearman correlation. The resulting method is described as semi-supervised, training-free, and was reported to be both efficient and highly competitive (Babnik et al., 22 Sep 2025).

5. Vision transformers, attention, and generic perceptual face IQA

Vision Transformer (ViT) architectures have become a distinct FIQA substrate. ViT-FIQA introduces a learnable quality token $NNCCS_{x_i} = \max_{j \ne y_i} x_i \cdot w_j,$9, concatenated with patch tokens and processed by global self-attention. The final backbone branches into a recognition head trained with CosFace and a regression head that predicts a scalar quality score from the quality token; the supervision target is the CR-FIQA score, optimized with Smooth L1 (Huber) loss, under the joint objective

$$CR_{x_i} = \frac{CCS_{x_i}}{NNCCS_{x_i} + (1 + \epsilon)}.$$0

The method was trained on MS1MV2 and reported top-tier performance across both CNN- and ViT-based FR backbones (Atzori et al., 19 Aug 2025).

Two later transformer methods removed FIQA training altogether. ATTN-FIQA uses pre-softmax attention matrices from the final transformer block of pre-trained ViT-based FR models. After flattening all heads, it defines the image-level score as the simple mean of all entries,

$$CR_{x_i} = \frac{CCS_{x_i}}{NNCCS_{x_i} + (1 + \epsilon)}.$$1

It requires only a single forward pass, no architectural modifications, and no additional training; the paper emphasizes that the attention maps provide spatial interpretability, revealing which facial regions contribute most to quality determination (Ozgur et al., 21 Apr 2026).

ViTNT-FIQA instead measures the stability of patch embedding evolution across consecutive transformer blocks. With normalized patch embeddings $$CR_{x_i} = \frac{CCS_{x_i}}{NNCCS_{x_i} + (1 + \epsilon)}.$$2, it computes Euclidean distances between blocks, averages them per patch, and converts the result to a patch quality score through

$$CR_{x_i} = \frac{CCS_{x_i}}{NNCCS_{x_i} + (1 + \epsilon)}.$$3

Image-level quality is then aggregated either uniformly or with attention weights from the last transformer block. The method is training-free, uses a single forward pass without backpropagation, and was validated both on a synthetic quality-labeled dataset and on eight standard FIQA benchmarks (Ozgur et al., 9 Jan 2026).

A parallel line addresses Generic Face Image Quality Assessment (GFIQA) rather than biometric utility. DSL-FIQA explicitly targets perceptual quality and couples a Dual-Set Degradation Representation Learning (DSL) mechanism with a Landmark-Guided Transformer. The DSL encoder is trained with synthetic and real degradations so as to decouple degradation from facial content, while the transformer uses landmark identifiers to emphasize visually salient regions. The method is trained to regress MOS with Charbonnier loss, and it is supported by CGFIQA-40k, a dataset of 39,312 images with 53.83% light, 24.91% medium, 21.26% dark, 51.50% male, and 48.50% female representation (Chen et al., 2024).

6. Bias, fairness, and interpretability

Bias analyses show that FIQA inherits and can amplify the biases of the underlying FR systems. On the Balanced Faces in the Wild (BFW) dataset, a comprehensive study found that all techniques considered are affected more by variations in race than sex; general-purpose image quality assessment methods appear to be less biased with respect to race and sex, while supervised and unsupervised face image quality assessment methods both show strong bias with a tendency to favor white individuals. The study further reports a trade-off: methods that are less racially biased perform worse overall. For ArcFace, CR-FIQA yielded $$CR_{x_i} = \frac{CCS_{x_i}}{NNCCS_{x_i} + (1 + \epsilon)}.$$4 for Whites and $$CR_{x_i} = \frac{CCS_{x_i}}{NNCCS_{x_i} + (1 + \epsilon)}.$$5 for Blacks (+40% relative), whereas NIQE yielded $$CR_{x_i} = \frac{CCS_{x_i}}{NNCCS_{x_i} + (1 + \epsilon)}.$$6 for Whites and $$CR_{x_i} = \frac{CCS_{x_i}}{NNCCS_{x_i} + (1 + \epsilon)}.$$7 for Blacks (+1.3%), illustrating lower bias but lower utility (Babnik et al., 2022).

One attempt to make FIQA more demographic-agnostic focuses on the eye sclera. Because the sclera has a consistent whitish color regardless of age, ethnicity, and skin-tone, a landmark-based sclera segmentation method was proposed for enrollment and border-control scenarios. Pixels inside the convex hull of the eye landmarks but outside the minimum enclosing circle of the iris are treated as sclera pixels, and simple statistics such as mean pixel value are shown to respond consistently to changes in saturation, white balance, and exposure across subjects with light and dark skin. The paper argues that these sclera-derived features can be incorporated into FIQA algorithms to make them agnostic to demographic factors (Kabbani et al., 2023).

Interpretability has also motivated zero-shot use of Vision-LLMs (VLMs). Off-the-shelf VLMs were evaluated as FIQA systems that can output both a scalar score and human-understandable justifications such as “Low Resolution” or “Blurry.” Their biometric utility depends significantly on architecture, not merely on parameter count; larger models are more internally consistent and less prompt-sensitive, but may hallucinate degradations more often on clean images. Because they are much slower than conventional FIQA, the study positions them as a promising interpretability module for human-in-the-loop scenarios rather than as a front-line high-throughput component (Sarıtaş et al., 17 May 2026).

7. Evaluation protocols, efficiency constraints, and open problems

Most biometric-utility FIQA studies evaluate quality ranking through Error-versus-Reject / Error-versus-Discard curves: the False Non-Match Rate (FNMR) is tracked at a fixed False Match Rate (FMR) as increasing fractions of low-quality images are rejected. Summary measures include AUC and partial AUC (pAUC), where lower values indicate better quality assessment. One formulation used in later work is

$$CR_{x_i} = \frac{CCS_{x_i}}{NNCCS_{x_i} + (1 + \epsilon)}.$$8

with $$CR_{x_i} = \frac{CCS_{x_i}}{NNCCS_{x_i} + (1 + \epsilon)}.$$9 the rejection ratio. Bias studies additionally use normalized $$Q_{x_i} = \mathbb{WD}\left(\mathcal{S}^P_{x_i} \,\|\, \mathcal{S}^N_{x_i}\right).$$0 to compare demographic groups independently of absolute FR performance (Babnik et al., 2022, Babnik et al., 2023, Sarıtaş et al., 17 May 2026).

Perceptual FIQA benchmarks increasingly rely on Spearman Rank-Order Correlation Coefficient (SROCC) and Pearson Linear Correlation Coefficient (PLCC) against MOS labels. The VQualA 2025 Challenge on Face Image Quality Assessment imposed strict efficiency constraints of 0.5 GFLOPs and 5 million parameters and ranked submissions by

$$Q_{x_i} = \mathbb{WD}\left(\mathcal{S}^P_{x_i} \,\|\, \mathcal{S}^N_{x_i}\right).$$1

The challenge report states that it attracted 127 participants and 1519 final submissions, with the top-ranked entry achieving 0.9664 under the lightweight constraint (Ma et al., 25 Aug 2025).

Efficiency has become a central research theme. DifFIQA’s original diffusion-based inference is computationally heavy at 1074 ms/img, whereas DifFIQA(R) reduces this to 1.24 ms/img, approximately 1000x faster, while maintaining competitive performance (Babnik et al., 2023). In the VQualA setting, a two-stage self-training and distillation framework trained a Swin Transformer Base teacher (86.88M params, 50.13 GFLOPs) and distilled an EdgeNeXt-XX-Small student (1.18M params, 0.4949 GFLOPs) that achieved SRCC = 0.9855 and PLCC = 0.9914, while the enhanced teacher achieved SRCC = 0.9867 and PLCC = 0.9926 (Sun et al., 21 Jul 2025). A separate lightweight ensemble of MobileNetV3-Small and ShuffleNetV2 with MSECorrLoss reported SRCC = 0.9829 and PLCC = 0.9894 on the VQualA benchmark while remaining within the competition efficiency constraints (Hamidi et al., 12 Sep 2025).

The survey literature identifies several unresolved issues: comparability and reproducibility, explainability/interpretability, training data and labeling, interoperability across FR systems, robustness/vulnerability, standardization, and extension to new spectra/modalities/tasks. This suggests that FIQA remains a moving target defined jointly by the target FR system, the operating constraints, the desired level of transparency, and the choice between biometric utility and perceptual facial quality as the optimization objective (Schlett et al., 2020).