Image Intrinsic Scale Assessment: Bridging the Gap Between Quality and Resolution

Daisuke Iso; Dietmar Saupe; Lorenzo Agnolucci; Vlad Hosu

REVIEW 2 major objections 4 minor 41 references

Image Intrinsic Scale Assessment: Bridging the Gap Between Quality and Resolution

T0 review · 2 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper introduces the Image Intrinsic Scale, the size at which an image looks best, and shows that weak labels derived from this scale improve automatic prediction across six quality-assessment models.

desk verdict The IIS task and IISA-DB dataset are the real contribution; the WIISA weak-label mechanism is plausible but under-validated and needs a control experiment. read the letter →

arxiv 2502.06476 v3 pith:XYSYXCLU submitted 2025-02-10 cs.CV

Vlad Hosu , Lorenzo Agnolucci , Daisuke Iso , Dietmar Saupe This is my paper · ORCID

classification cs.CV

keywords imageintrinsicscaleperceptualqualityno-referenceassessmentweaklabelingdownscalingcrowdsourcedannotationresolutionIISA

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to quantify something familiar: every photo, when shrunk, first looks cleaner and then loses detail, so there is a middle scale where it looks best. It calls that best size the Image Intrinsic Scale (IIS), defines the new task of predicting it, and collects 785 expert-annotated image–IIS pairs in the IISA-DB dataset. The paper then proposes WIISA, a rule that derives weak labels for downscaled copies from a single ground-truth label, and reports that adding these weak labels improves prediction for every one of the six no-reference quality models tested. If these results hold, the IIS becomes a practical benchmark for scale-aware quality, and any quality model can gain accuracy by training with the weak-label rule.

What carries the argument

The load-bearing construction is the piecewise extrapolation identity of Eq. (2). It transforms one human judgment—the intrinsic scale of an image at full resolution—into weak labels for all downscaled versions: a version already at or below the intrinsic scale keeps intrinsic scale 1, while a version above it has intrinsic scale equal to $\Omega(I)/s$. The other central element is the annotation tool, a slider that lets experts shrink an image until quality stops improving, with labels aggregated through the geometric mean to form the Mean Opinion Intrinsic Scale. The identity is what makes weak labeling cheap and model-agnostic: any regression head or quality network can consume the generated pairs.

What would settle it

Run a fine-grained subjective study with many more scale levels (e.g., 20 or more) on a diverse set of images and look at each image's quality-versus-scale curve: if a substantial fraction of curves show two separated peaks or a re-rise after a dip, the single-peak assumption behind Eq. (2) fails, and WIISA's weak labels become wrong for those images.

Watch

Extended reading notes

Core claim

The central discovery the paper argues for is that perceived image quality, as a function of downscaling factor, is shaped like a single hill: it rises as compression artifacts, noise, and blur become less visible, then falls once meaningful high-frequency detail is lost. Because of this unimodal shape, each image has a unique intrinsic scale $\Omega(I)$, defined as the largest scale at which its quality peaks. The paper further claims that the intrinsic scale of a downscaled version is fully determined by the original one through the rule $\Omega(I^s)=1$ when $s\leq\Omega(I)$ and $\Omega(I^s)=\Omega(I)/s$ when $s>\Omega(I)$. Training with labels generated by this rule, on top of ground-truth labels, is reported to improve Spearman and Pearson correlations for six NR-IQA architectures adapted to the task, with the best model reaching SRCC 0.808 on the new benchmark.

Load-bearing premise

The whole scheme rests on the assumption that as an image is progressively downscaled its perceived quality first climbs and then falls—one peak, no wiggles—so the image has a single best scale.

Editorial extensions

If this is right

IISA-DB provides the first benchmark where models are scored by how well they predict the scale at which an image looks best, enabling direct comparison of scale-aware quality methods.
Because the weak-label rule costs only a downscale operation, training any IISA model with WIISA becomes a drop-in improvement for architectures from supervised CNNs to vision-language models.
The IIS gives a concrete content-aware rule for display and storage: downscale an image to its intrinsic scale to maximize perceived quality while saving bandwidth, print area, or rendering time.
The sensitivity analysis in the supplementary material implies that IISA can detect subtle quality differences—such as those from denoising or super-resolution—that ordinary mean opinion scores do not reliably resolve.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper, the same weak-label rule could be used to propagate IIS annotations to synthetically distorted or super-resolved images, turning one label into labels for many distortions.
Going beyond the paper, a fine-grained scale-sweep study would likely reveal that the single-peak assumption holds less cleanly for images with strong texture or repeated patterns, where downscaling can create moiré and aliasing artifacts that reintroduce perceived defects.
The paper's logarithm-of-IIS versus MOS linear relationship suggests that scale labels could be reinterpreted as a perceptual resolution axis, potentially connecting IISA to effective-resolution and just-noticeable-difference tasks.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Dear colleague,

Here is my take on arXiv:2502.06476. The real news is IISA-DB: a new perceptual quantity (IIS), a carefully collected 785-image benchmark with expert annotations, and a clean task definition that clearly separates the subjective and predictive sides. The dataset methodology is solid — 20 opinions per image, repeated labeling, quality-control re-annotation, geometric-mean aggregation on a log slider, and confidence intervals in the same range as KonX. The release of code, data, and pretrained models makes the empirical claims checkable.

The WIISA weak-labeling idea is elegant and the gains in Table 2 are consistent across six different NR-IQA backbones. But I am not convinced the mechanism is the one the paper proposes. The load-bearing assumption is monotonic quality increase below IIS and monotonic decrease above it. That is verified on KonX at three discrete resolutions, and 10% of those triplets already break the assumed shape. Worse, the weak labels are generated in the range [max(Omega, 0.65), 1], where only the second branch of Eq. 2 applies, and this is exactly the branch that is tested with only one or two discrete steps. There is no direct evidence that Omega(I^s) = Omega(I)/s on the continuum. The paper attributes the 10% violations to annotation noise, but the resampling analysis in S1.3 only shows that no single image can be confidently called a violation — it does not show that the assumption holds with confidence.

The more basic missing experiment is a control: train with weak labels generated by a deliberately wrong rule, or with the same number of extra downscaled images but random labels. Without that, I cannot tell whether the boosted performance reflects genuine IIS-specific structure or simply the regularizing effect of more training samples. The reported gains are also modest (up to 5%, some under 1 point of SRCC) and the paper reports medians over 10 splits without error bars, so part of the effect might be within fold noise.

That said, I would not block the paper on this. The dataset alone is a genuine contribution, and the concept of IIS is useful for display, printing, and dataset curation. The weak-label story needs more evidence, but the authors are transparent about the assumption and the limitations section is honest. The citation pattern is fine — differences with perceptual megapixels and effective resolution are clearly drawn. This paper deserves a serious referee; the revision should ask for the control experiment, error bars, and finer-grained validation of the monotonicity assumption outside KonX.

Who should read it: anyone working on IQA, computational photography, or dataset curation. I would bring it to our reading group.

Recommendation: accept into peer review, expect moderate revision.

Referee Report

2 major / 4 minor

Summary. The paper introduces the Image Intrinsic Scale (IIS), the largest scale at which an image shows its highest perceived quality, and defines the Image Intrinsic Scale Assessment (IISA) task. The authors collect the IISA-DB dataset of 785 image-IIS pairs through a carefully controlled expert crowdsourcing study with repeated annotations and reliability checks. They then propose WIISA, a weak-labeling strategy that uses a piecewise extrapolation rule (Eq. 2) to generate additional IIS labels for downscaled versions of each image. Experiments report that applying WIISA during training consistently improves the performance of six NR-IQA methods adapted for IISA, with gains up to about 5% in SRCC.

Significance. The paper makes a novel contribution by defining and operationalizing a perceptual scale-assessment task with a publicly released dataset, code, and pre-trained models. The annotation methodology is rigorous: 20 opinions per image from expert freelancers, repeated labeling, re-annotation below an SRCC threshold, and confidence intervals comparable to the KonX dataset. The evaluation is not circular: ground-truth labels are human judgments, while WIISA generates weak labels from those ground truths via Eq. 2 without fitting to test labels. If the weak-labeling rule is valid, WIISA offers a model-agnostic, low-cost boost for any IISA predictor, and IISA could become a useful benchmark for scale-aware quality assessment. The main risk to the central claim is the insufficient empirical validation of the monotonicity assumptions A1/A2 on which Eq. 2 is based, and the lack of significance testing for the reported improvements.

major comments (2)

[Sec. 6.3 and S1.3; Eq. (2)] The load-bearing assumption A1/A2 (Sec. 3) is validated only on the KonX dataset, which provides MOS at three resolutions. The paper reports that 378/420 triplets are consistent with the assumed shape, but the 42 remaining triplets violate it, and the resampling analysis in S1.3 only shows that for no single image the violation is statistically certain at the 90% level; it does not demonstrate that the violations are annotation noise. Since WIISA samples scales in [max(Omega(I), delta), 1] (Sec. 5.2), it always uses the second branch of Eq. (2), which relies on monotonic decrease of Q(I^s) for s > Omega(I). This monotonicity is not tested at continuous scales or on the IISA-DB image distribution. Consequently, the weak labels for a non-negligible fraction of training images may be systematically wrong, and the consistent gains in Table 2 could be due to the additional training data acting as augmentation or regularization rather than to the correctness of the IIS-specific weak-labeling rule. Please add a control experiment that separates the effect of extra data from the correctness of Eq. 2 (e.g., random or corrupted weak labels, or downscaled images labeled with a different rule), and/or validate Eq. 2 on a subset of IISA-DB by collecting subjective IIS annotations on downscaled versions of the same images.
[Table 2 and Sec. 6.2] The paper reports median performance over 10 random splits but does not provide variance or significance tests. For several methods the WIISA gains are small (e.g., CLIP-IQA+ SRCC 0.660 to 0.666; QualiCLIP SRCC 0.665 to 0.670; MAE changes are also small). Without confidence intervals or paired significance tests, it is unclear whether these differences are beyond cross-validation noise. Because the central claim is that WIISA 'consistently improves' performance, please report the distribution of metrics across the 10 splits and perform a paired significance test (e.g., Wilcoxon signed-rank) across splits, or at least provide per-split results for all methods.

minor comments (4)

[Eq. (S1) in S2.3] The geometric-mean formula appears typeset incorrectly (the exponent shows 'PN' instead of a summation/product notation). Please fix the LaTeX so that the formula reads as the geometric mean of the 20 opinions.
[Sec. S3.1] The text says 'we extract square center crops with a size of 1536 pixels.' When a downscaled image has width smaller than 1536 pixels, it is unclear whether the image is resized or padded. Please clarify the procedure for such cases.
[Sec. 3] The sentence 'In the supplementary material, we prove that IISA significantly outperforms traditional NR-IQA quality ratings in detecting subtle differences' overstates what S1.1 shows: the leverage analysis is an indirect argument based on KonX aggregate MOS, not a proof of IISA sensitivity. Please soften the wording.
[Sec. 6.3] The discussion of the delta threshold says lower delta values produce weak labels that may be 'out of the ground-truth distribution'; please specify the distribution concretely (e.g., IISA-DB MOIS mean 0.347, range 0.060-0.811) to make the claim quantitative.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: WIISA's weak labels are derived from explicit assumptions and tested against independent human ground-truth labels.

full rationale

The derivation chain is self-contained. Equation (1) defines IIS as a theoretical argmax over the latent quality curve Q(I^s), and Equation (2) is not a fitted input-output rule; it is derived from the stated assumptions A1/A2. WIISA generates weak labels from ground-truth human labels using that formula, and the central improvement claim in Table 2 is evaluated under a standard train/validation/test protocol against independent ground-truth MOIS labels from IISA-DB. Weak labels are never used to define the test target, and the hyperparameters delta and nwl are selected based on validation behavior rather than fitted to test labels. The KonX [34] validation of A1/A2 is an external empirical dataset with its own subjective MOS values, not a uniqueness theorem or a fitted parameter imported from the present authors; it is corroborating evidence, and the assumption is also supported by independent viewing-distance studies [8, 11, 25]. No step reduces by construction to its own input: the weak-label formula is not the ground-truth definition, the trained models do not see test weak labels, and no claim of uniqueness or forced choice is made. Concerns about the assumption being validated only at three discrete KonX scales, and about the 10% of triplets inconsistent with the assumed shape, are empirical robustness issues rather than definitional circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 1 invented entities

The central claim depends on the monotonicity assumption for quality versus scale, on the aggregation of noisy subjective opinions into MOIS, and on the choice of interpolation. The weak-label formula in Eq. 2 is not fitted to the test labels, so the circularity burden is low, but the continuous extrapolation of the monotonicity assumption is the least externally supported link.

free parameters (4)

s_lb (lower bound for scale) = 0.05
Chosen in Sec. 3 as the minimum scale for IIS measurement; it defines the search range in Eq. 1 and affects all IIS labels in IISA-DB.
delta (weak-label downscaling threshold) = 0.65
Empirically set in Sec. 5.2; it controls the sampled scale range [max(Omega(I), delta), 1] for WIISA, and varying it changes results in Table 3.
n_wl (number of weak labels per image) = 2
Selected in Sec. 5.2; Table 3 shows performance peaks at n_wl=2 and declines for n_wl=3, so this is a tuned hyperparameter.
n_s (scales sampled in zero-shot multi-scale baseline) = 100
Used only for the zero-shot multi-scale IISA baseline in Table S1, not for the main WIISA results; included for completeness.

assumptions (3)

domain assumption A1/A2: perceived quality Q(I^s) increases monotonically for s <= Omega(I) and decreases monotonically for s > Omega(I).
Sec. 3 states this assumption and uses it to derive Eq. 2. It is validated on the KonX dataset with 378 of 420 images consistent, but only three discrete resolutions are sampled.
domain assumption The geometric mean of 20 expert IIS annotations approximates the latent expected IIS for each image.
Sec. 4.1 and Sec. S2.3 justify the MOIS aggregation by the non-linear slider scale; this relies on the representativeness of the expert annotator pool.
domain assumption Lanczos interpolation preserves the perceived quality-scale relationship relevant to IIS.
Sec. 3 and Sec. S4 use Lanczos for both subjective annotation and weak-label generation; the paper acknowledges other interpolation methods could change artifacts, though the predictive ablation in Table 3 shows limited impact.

invented entities (1)

Image Intrinsic Scale (IIS) independent evidence
purpose: Central new quantity: the largest scale at which an image shows its highest perceived quality, enabling a new prediction task.
Defined in Sec. 3 Eq. 1 and directly measured by human annotators in the IISA-DB dataset, giving it a falsifiable measurement handle outside the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Image Intrinsic Scale Assessment: Bridging the Gap Between Quality and Resolution." pith.science (2026). https://pith.science/paper/XYSYXCLU

@misc{pith2026250206476,
  author       = {Pith},
  title        = {Pith review of: Image Intrinsic Scale Assessment: Bridging the Gap Between Quality and Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XYSYXCLU}},
  note         = {Machine review of arXiv:2502.06476}
}

read the original abstract

Image Quality Assessment (IQA) measures and predicts perceived image quality by human observers. Although recent studies have highlighted the critical influence that variations in the scale of an image have on its perceived quality, this relationship has not been systematically quantified. To bridge this gap, we introduce the Image Intrinsic Scale (IIS), defined as the largest scale where an image exhibits its highest perceived quality. We also present the Image Intrinsic Scale Assessment (IISA) task, which involves subjectively measuring and predicting the IIS based on human judgments. We develop a subjective annotation methodology and create the IISA-DB dataset, comprising 785 image-IIS pairs annotated by experts in a rigorously controlled crowdsourcing study. Furthermore, we propose WIISA (Weak-labeling for Image Intrinsic Scale Assessment), a strategy that leverages how the IIS of an image varies with downscaling to generate weak labels. Experiments show that applying WIISA during the training of several IQA methods adapted for IISA consistently improves the performance compared to using only ground-truth labels. The code, dataset, and pre-trained models are available at https://github.com/SonyResearch/IISA.

Figures

Figures reproduced from arXiv: 2502.06476 by the authors.

**Figure 1.** Top: downscaling an image affects its perceived quality. As the original image (left) is downscaled, degradation becomes less noticeable, but some high-frequency details may be lost. The optimal trade-off occurs at the image’s intrinsic scale. Note that the first two images are cropped for clearer visualization. Bottom: we plot several possible profiles corresponding to different images, depicting how the quality ch… view at source ↗

**Figure 2.** Example of the influence of the scaling factor [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Comparison of inter-group agreement between the pro [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Overview of WIISA, the proposed weak-label gener [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 38 canonical work pages

[1]

Quality-Aware Image-Text Alignment for Opinion- Unaware Image Quality Assessment

Lorenzo Agnolucci, Leonardo Galteri, and Marco Bertini. Quality-Aware Image-Text Alignment for Opinion- Unaware Image Quality Assessment. arXiv preprint arXiv:2403.11176, 2024. 3, 8

arXiv 2024
[2]

ARNIQA: Learning Distortion Mani- fold for Image Quality Assessment

Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. ARNIQA: Learning Distortion Mani- fold for Image Quality Assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 189–198, 2024. 3, 8

work page 2024
[3]

Glenn D. Boreman. Modulation Transfer Function in Opti- cal and Electro-Optical Systems . SPIE Press, Bellingham, W A, 2001. 3

work page 2001
[4]

Intrinsic Camera Resolution Measurement

Peter D Burns and Judit Martinez Bauza. Intrinsic Camera Resolution Measurement. In Image Quality and System Per- formance XII, pages 55–67. SPIE, 2015. 3

work page 2015
[5]

TOPIQ: A Top-down Approach from Semantics to Distortions for Im- age Quality Assessment

Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. TOPIQ: A Top-down Approach from Semantics to Distortions for Im- age Quality Assessment. IEEE Transactions on Image Pro- cessing, 2024. 3, 6, 7, 8, 14, 15

work page 2024
[6]

Scale Develop- ment: Theory and Applications

Robert F DeVellis and Carolyn T Thorpe. Scale Develop- ment: Theory and Applications. Sage Publications, 2021. 4, 12

work page 2021
[7]

Perceptual MegaPixel

DxOMark. Perceptual MegaPixel. https : / / en . wikipedia . org / w / index . php ? title = Perceptual _ MegaPixel & oldid = 878646818 ,

work page
[8]

Evaluation of Image Quality of Experience in Consideration of Viewing Distance

Ruigang Fang, Dapeng Wu, and Liquan Shen. Evaluation of Image Quality of Experience in Consideration of Viewing Distance. In IEEE China Summit and International Con- ference on Signal and Information Processing (ChinaSIP) , pages 653–657. IEEE, 2015. 3, 8, 12

work page 2015

Show all 41 references

[9]

Perceptual Quality Assessment of Smartphone Pho- tography

Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual Quality Assessment of Smartphone Pho- tography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 3677–3686, 2020. 2, 5, 6, 15

2020
[10]

No-reference Image Quality Assessment via Transformers, Relative Ranking, and Self-consistency

S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference Image Quality Assessment via Transformers, Relative Ranking, and Self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, pages 1220–1230, 2022. 3

2022
[11]

Quality Assessment Considering Viewing Dis- tance and Image Resolution

Ke Gu, Min Liu, Guangtao Zhai, Xiaokang Yang, and Wen- jun Zhang. Quality Assessment Considering Viewing Dis- tance and Image Resolution. IEEE Transactions on Broad- casting (ToB), 61(3):520–531, 2015. 3, 8, 12

2015
[12]

DIV8K: DI- Verse 8K Resolution Image Dataset

Shuhang Gu, Andreas Lugmayr, Martin Danelljan, Manuel Fritsche, Julien Lamour, and Radu Timofte. DIV8K: DI- Verse 8K Resolution Image Dataset. In IEEE/CVF Interna- tional Conference on Computer Vision Workshop (ICCVW), pages 3512–3516. IEEE, 2019. 5

2019
[13]

The Effect of Viewing Distance and Dis- play Peak Luminance - HDR A V1 Video Streaming Quality Dataset

Dounia Hammou, Luk ´aˇs Krasula, Christos Bampis, Zhi Li, and Rafal Mantiuk. The Effect of Viewing Distance and Dis- play Peak Luminance - HDR A V1 Video Streaming Quality Dataset. In Proceedings of the International Conference on Quality of Multimedia Experience (QoMEX), 2024. 3, 4

2024
[14]

KonIQ-10k: an Ecologically Valid Database for Deep Learn- ing of Blind Image Quality Assessment

Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. KonIQ-10k: an Ecologically Valid Database for Deep Learn- ing of Blind Image Quality Assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 2, 5, 6, 11, 13

2020
[15]

UHD-IQA Benchmark Database: Pushing the Boundaries of Blind Photo Quality Assessment

Vlad Hosu, Lorenzo Agnolucci, Oliver Wiedemann, and Daisuke Iso. UHD-IQA Benchmark Database: Pushing the Boundaries of Blind Photo Quality Assessment. arXiv preprint arXiv:2406.17472, 2024. 2, 5, 6, 15

2024 arXiv
[16]

High Resolu- tion Image Quality Database

Huang Huang, Qiang Wan, and Jari Korhonen. High Resolu- tion Image Quality Database. In IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pages 3105–3109. IEEE, 2024. 15

2024
[17]

ITUT. Recommendation ITUT P.913 (06/2021) Methods for the Subjective Assessment of Video Quality, Audio Quality and Audiovisual Quality of Internet Video and Distribution Quality Television in Any Environment, 2021. 5

2021
[18]

Statistical Study on Perceived JPEG Image Quality via MCL-JCI Dataset Construction and Analysis

Lina Jin, Joe Yuchieh Lin, Sudeng Hu, Haiqiang Wang, Ping Wang, Ioannis Katsavounidis, Anne Aaron, and C-C Jay Kuo. Statistical Study on Perceived JPEG Image Quality via MCL-JCI Dataset Construction and Analysis. Electronic Imaging, 2016(13):1–9, 2016. 5

2016
[19]

Self-Supervised Effective Resolution Esti- mation with Adversarial Augmentations

Manuel Kansy, Julian Balletshofer, Jacek Naruniec, Christo- pher Schroers, Graziana Mignone, Markus Gross, and Ro- mann M Weber. Self-Supervised Effective Resolution Esti- mation with Adversarial Augmentations. In Proceedings of the IEEE/CVF Winter Conference on Applications o...

2023
[20]

MUSIQ: Multi-scale Image Quality Trans- former

Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale Image Quality Trans- former. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5148–5157,
[21]

Influence of Viewing Distances on 8K HDR Video Quality Perception

Dominik Keller, Felix von Hagen, Julius Prenzel, Kay Strama, Rakesh Rao, Ramachandra Rao, and Alexander Raake. Influence of Viewing Distances on 8K HDR Video Quality Perception. In International Conference on Quality of Multimedia Experience (QoMEX), pages 209–212. IEEE,
[22]

Visual Quality Assessment Considering Ultra HD, Full HD Resolution and Viewing Dis- tance

Jan Kufa and Tomas Kratochvil. Visual Quality Assessment Considering Ultra HD, Full HD Resolution and Viewing Dis- tance. In International Conference Radioelektronika (RA- DIOELEKTRONIKA), pages 1–4. IEEE, 2019. 3

2019
[23]

LSDIR: A Large Scale Dataset for Image Restora- tion

Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Deman- dolx, et al. LSDIR: A Large Scale Dataset for Image Restora- tion. In IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 1775–1787, 2023. 5

2023
[24]

Large-scale Crowdsourced Subjective Assessment of Pic- turewise Just Noticeable Difference

Hanhe Lin, Guangan Chen, Mohsen Jenadeleh, Vlad Hosu, Ulf-Dietrich Reips, Raouf Hamzaoui, and Dietmar Saupe. Large-scale Crowdsourced Subjective Assessment of Pic- turewise Just Noticeable Difference. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5859– ...

2022
[25]

CID: IQ - a New Image Quality Database

Xinwei Liu, Marius Pedersen, and Jon Yngve Hardeberg. CID: IQ - a New Image Quality Database. In International Conference on Image Signal Processing (ICISP), pages 193–
[26]

Image Quality Assessment Using Contrastive Learning

Pavan C Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C Bovik. Image Quality Assessment Using Contrastive Learning. IEEE Transactions on Image Processing, 31:4149–4161, 2022. 3, 8

2022
[27]

Learning Transferable Visual Models from Natural Language Super- vision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models from Natural Language Super- vision. In International Conference on Machine Learning ,...

2021
[28]

Re- IQA: Unsupervised Learning for Image Quality Assessment in the Wild

Avinab Saha, Sandeep Mishra, and Alan C Bovik. Re- IQA: Unsupervised Learning for Image Quality Assessment in the Wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5846– 5855, 2023. 3

2023
[29]

Real Versus Fake 4K-authentic Resolution Assessment

Rishi Rajesh Shah, Vyas Anirudh Akundy, and Zhou Wang. Real Versus Fake 4K-authentic Resolution Assessment. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2185–2189. IEEE, 2021. 3

2021
[30]

A JND dataset based on VVC compressed images

Xuelin Shen, Zhangkai Ni, Wenhan Yang, Xinfeng Zhang, Shiqi Wang, and Sam Kwong. A JND dataset based on VVC compressed images. In 2020 IEEE International Con- ference on Multimedia & Expo Workshops (ICMEW), pages 1–6. IEEE, 2020. 5

2020
[31]

Blindly Assess Image Qual- ity in the Wild Guided by a Self-adaptive Hyper Network

Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly Assess Image Qual- ity in the Wild Guided by a Self-adaptive Hyper Network. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 3667–3676, 2020. 3

2020
[32]

JPEG AIC-3 Dataset: Towards Defining the High Quality to Nearly Visu- ally Lossless Quality Range

Michela Testolina, Vlad Hosu, Mohsen Jenadeleh, Davi Laz- zarotto, Dietmar Saupe, and Touradj Ebrahimi. JPEG AIC-3 Dataset: Towards Defining the High Quality to Nearly Visu- ally Lossless Quality Range. In International Conference on Quality of Multimedia Experience (QoMEX) , ...

2023
[33]

Ex- ploring CLIP for Assessing the Look and Feel of Images

Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Ex- ploring CLIP for Assessing the Look and Feel of Images. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 2555–2563, 2023. 3, 8

2023
[34]

KonX: Cross-Resolution Image Quality Assessment

Oliver Wiedemann, Vlad Hosu, Shaolin Su, and Dietmar Saupe. KonX: Cross-Resolution Image Quality Assessment. Quality and User Experience, 8(1):8, 2023. 2, 3, 4, 6, 8, 11, 12, 13, 15

2023
[35]

Blind Image Quality Assessment Using a Deep Bi- linear Convolutional Neural Network

Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind Image Quality Assessment Using a Deep Bi- linear Convolutional Neural Network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36–47,
[36]

Blind Image Quality Assessment via Vision- Language Correspondence: A Multitask Learning Perspec- tive

Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind Image Quality Assessment via Vision- Language Correspondence: A Multitask Learning Perspec- tive. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 14071–14081,
[37]

Quality-Aware Pre-Trained Models for Blind Image Quality Assessment

Kai Zhao, Kun Yuan, Ming Sun, Mading Li, and Xing Wen. Quality-Aware Pre-Trained Models for Blind Image Quality Assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22302– 22313, 2023. 3

2023
[38]

continuous

Wenhan Zhu, Guangtao Zhai, Xiongkuo Min, Xiaokang Yang, and Xiao-Ping Zhang. Perceptual Quality Assessment for Recognizing True and Pseudo 4K Content. In IEEE In- ternational Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), pages 2190–2194. IEEE, 2021. 3 10 Im...

2021
[41]

( ·) indicates the pre-training dataset

model. ( ·) indicates the pre-training dataset. Best scores are highlighted in bold. approach on the proposed IISA-DB dataset. We employ two versions of the TOPIQ [5] model pre-trained on the SPAQ [9] and UHD-IQA [15] datasets, which feature high- resolution images similar to ...
[202]

3, 8, 12 9

Springer, 2014. 3, 8, 12 9

2014
[2012]

[Online; accessed 2024/08/05]. 3

2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Quality-Aware Image-Text Alignment for Opinion- Unaware Image Quality Assessment

Lorenzo Agnolucci, Leonardo Galteri, and Marco Bertini. Quality-Aware Image-Text Alignment for Opinion- Unaware Image Quality Assessment. arXiv preprint arXiv:2403.11176, 2024. 3, 8

arXiv 2024

[2] [2]

ARNIQA: Learning Distortion Mani- fold for Image Quality Assessment

Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. ARNIQA: Learning Distortion Mani- fold for Image Quality Assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 189–198, 2024. 3, 8

work page 2024

[3] [3]

Glenn D. Boreman. Modulation Transfer Function in Opti- cal and Electro-Optical Systems . SPIE Press, Bellingham, W A, 2001. 3

work page 2001

[4] [4]

Intrinsic Camera Resolution Measurement

Peter D Burns and Judit Martinez Bauza. Intrinsic Camera Resolution Measurement. In Image Quality and System Per- formance XII, pages 55–67. SPIE, 2015. 3

work page 2015

[5] [5]

TOPIQ: A Top-down Approach from Semantics to Distortions for Im- age Quality Assessment

Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. TOPIQ: A Top-down Approach from Semantics to Distortions for Im- age Quality Assessment. IEEE Transactions on Image Pro- cessing, 2024. 3, 6, 7, 8, 14, 15

work page 2024

[6] [6]

Scale Develop- ment: Theory and Applications

Robert F DeVellis and Carolyn T Thorpe. Scale Develop- ment: Theory and Applications. Sage Publications, 2021. 4, 12

work page 2021

[7] [7]

Perceptual MegaPixel

DxOMark. Perceptual MegaPixel. https : / / en . wikipedia . org / w / index . php ? title = Perceptual _ MegaPixel & oldid = 878646818 ,

work page

[8] [8]

Evaluation of Image Quality of Experience in Consideration of Viewing Distance

Ruigang Fang, Dapeng Wu, and Liquan Shen. Evaluation of Image Quality of Experience in Consideration of Viewing Distance. In IEEE China Summit and International Con- ference on Signal and Information Processing (ChinaSIP) , pages 653–657. IEEE, 2015. 3, 8, 12

work page 2015