Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

BPCLIP: A Bottom-up Image Quality Assessment from Distortion to Semantics Based on CLIP

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read BPCLIP claims that propagating distortion information from low to high levels with a frozen CLIP text encoder yields state-of-the-art image quality scores.

desk verdict Competent incremental IQA architecture whose headline gains over TOPIQ mostly sit inside the noise; the missing prompt list and unpaired baselines keep it from being verifiable now. read the letter →

arxiv 2506.17969 v1 pith:L3PKZFCV submitted 2025-06-22 cs.CV

classification cs.CV
keywords imagequalityassessmentCLIPbottom-upfeaturepropagationmultiscalecrossattentionfull-referenceIQAno-referencetext-imagealignmentperceptual
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 tries to establish that image quality is best assessed by tracking how low-level distortions disrupt high-level semantic features, rather than fusing multiscale features in parallel or guiding low levels from the top down. Its BPCLIP model propagates features bottom-up through cross-attention blocks and then scores an image by the cosine similarity between those features and text embeddings of 40 quality adjectives, with the CLIP text encoder kept frozen. On the four full-reference and three no-reference benchmarks the paper tests, it reports correlation coefficients at or near the top of the field while using a ResNet50 backbone and roughly half the inference cost of transformer-based rivals. If the claim holds, accurate and partly explainable quality scores can be obtained from a simple architecture with a fixed language side.

What carries the argument

The load-bearing component is the multiscale cross-attention (MSCA) block: for adjacent feature levels $G_i$ and $G_{i+1}$, it forms the query from $G_i$ and the key and value from $G_{i+1}$, so low-level distortion information is progressively injected into semantic features. Gated local pooling normalizes all levels to a common spatial size, a parallel self-attention weight branch emphasizes semantically important regions, and a frozen CLIP text encoder supplies the embeddings of 40 quality adjectives; the cosine similarities between image and text features at each level are concatenated and passed through an MLP that regresses to mean opinion score.

What would settle it

Retrain BPCLIP under the same protocol but replace the 40 adjectives with a synonym set or with 40 random adjectives; if PLCC and SRCC change by only the reported standard-deviation margins, then the language anchor is not carrying the result and the claimed connection to human language is incidental.

Watch

Extended reading notes

Core claim

The central claim is that the direction of feature flow, not simply the number of scales, is what limits current IQA models. BPCLIP uses a bottom-up encoder in which each multiscale cross-attention block takes a lower-level feature map as query and a higher-level map as key and value, so distortion evidence from fine details is carried upward into semantically richer layers. At the top, the score is regressed from the cosine similarities between the hierarchical image features and fixed text embeddings of 40 quality adjectives spanning six dimensions, making the language part of the model interpretable and frozen. The paper reports that this design outperforms or matches existing methods on LIVE, CSIQ, TID2013, KADID-10k, CLIVE, KonIQ-10k, and SPAQ, and that the cross-dataset results stay strong.

Load-bearing premise

The 40 image-quality adjectives, which are never listed in the paper, are assumed to give the frozen CLIP text encoder a stable and complete enough vocabulary that cosine similarities between those fixed text embeddings and learned image features regress reliably to human opinion scores on every benchmark.

Editorial extensions

If this is right

  • State-of-the-art or competitive full-reference and no-reference scores can be reached with a ResNet50 backbone and a frozen text encoder, so the metric stays comparatively cheap to train and run.
  • Because the text encoder is fixed, the quality vocabulary provides a built-in explanation channel: which of the six adjective dimensions an image matches can be read off from the per-level similarities.
  • The same architecture covers both full-reference and no-reference settings by changing only the gated-pooling input, so a single model family can serve both tasks.
  • Cross-dataset results reported in the paper indicate that bottom-up propagation generalizes beyond the training distribution better than several top-down and transformer-based baselines.

Reading between the lines

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

  • The 40 adjectives are never enumerated, so a natural extension is to publish the exact prompts and test whether a synonym set or a randomly selected adjective set preserves the reported gains; that would isolate whether the language anchor or the extra regression capacity does the work.
  • The bottom-up cross-attention design could transfer to other vision tasks where low-level artifacts alter semantic content, such as image restoration validation or quality assessment of generated images.
  • Because the text side is frozen, the method could in principle be adapted to new quality dimensions by editing the adjective list without retraining the encoder, though the paper does not test zero-shot or few-shot variants.
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.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes BPCLIP, a full-reference and no-reference image quality assessment model built on a CLIP image-text backbone. The image encoder uses a ResNet50 backbone with gated local pooling, a bottom-up multiscale cross-attention (MSCA) branch that propagates low-level distortion information to higher-level semantic features, and a self-attention weight branch. Quality scores are obtained by regressing on the cosine similarities between the resulting image features and frozen CLIP text embeddings of 40 claimed quality adjectives organized into six dimensions. The paper reports within-dataset FR results on LIVE, CSIQ, TID2013, and KADID-10k; cross-dataset FR results; NR results on CLIVE, KonIQ-10k, and SPAQ; an official-split KonIQ-10k check; ablations; and complexity comparisons. The headline claim is state-of-the-art or competitive performance with greater robustness and efficiency.

Significance. The architectural idea of propagating distortion information bottom-up rather than top-down is sensible and potentially valuable for IQA, and the use of a frozen CLIP text encoder with quality adjectives is an interesting step toward language-grounded quality assessment. The official-split KonIQ-10k result (PLCC 0.945 vs TOPIQ 0.939 in Table V) is a useful robustness check that goes beyond a single random-split protocol. However, the empirical support for the central 'superior results' claim is currently fragile: most headline margins over TOPIQ are within one reported standard deviation, and the paper does not establish that baseline numbers were obtained under matched per-split conditions. The missing enumeration of the 40 adjectives is a separate reproducibility gap that directly affects the method's core design. If the authors supply the prompt list and paired per-split statistics, the contribution could be solid; as written, the claims outrun the evidence.

major comments (4)
  1. [Section III.A; Tables II and IV] The paper states in Section III.A that it performs random splits 10 times and reports the mean and standard deviation, but it does not state whether the comparison methods in Tables II and IV were re-run under exactly the same splits or are quoted from prior papers. If the latter, the comparison confounds method with split protocol. The observed PLCC gaps over TOPIQ are LIVE +0.003, CSIQ +0.003, TID2013 +0.013, KADID-10k +0.004, CLIVE +0.003, KonIQ-10k +0.004, and SPAQ +0.006; for the first five of these the gap is no larger than BPCLIP's own reported standard deviation (e.g., ±0.003, ±0.003, ±0.013, ±0.014, and ±0.014, respectively). Please report paired per-split differences with means and standard deviations of the differences, or explicitly document that all methods were evaluated under identical splits.
  2. [Table VI; Section III.E] The ablation credits the bottom-up mechanism with +0.005 PLCC on KADID-10k (0.960 to 0.964) and the CLIP text encoder with +0.004 PLCC (0.960 to 0.964), yet the reported standard deviation for BPCLIP on KADID-10k in Table II is ±0.014. Without ablation-specific standard deviations or paired tests, the component-level claims in Section III.E are not quantitatively supported. Please report the same 10-split protocol, including means and standard deviations, for each ablation row.
  3. [Section II.A; Section IV] The 40 image quality adjectives and their six dimensions are referenced repeatedly but never enumerated, sourced, or justified. The reader cannot check whether the prompt set is benchmark-specific or was tuned with target-dataset MOS in the loop, and the missing list prevents replication of the method. Please provide the full list of prompts, the six dimensions, and, if any selection process was used, state whether it was done before or after seeing evaluation data.
  4. [Section III.C.2; Table III] Cross-dataset generalization is presented as evidence of 'greater robustness,' but Table III reports no standard deviations or significance tests, and the gaps over TOPIQ are small (PLCC 0.961 vs 0.955, 0.965 vs 0.962, and 0.922 vs 0.916). Please report per-split results or at least state how many runs were performed and whether the same protocol was used for the baseline methods.
minor comments (6)
  1. [Table II] The std row is inconsistently formatted: the entries '0.014 0.013' lack the ± symbols used elsewhere in the row.
  2. [Table VI] The symbols '#' and '!' used in the table rows are not defined in the caption or in Section III.E, making the ablation rows hard to interpret.
  3. [Section II.B, Eq. (2)] The linear projection W_f is introduced in Eq. (2) but is not defined or described in the text.
  4. [Section II.C, Eq. (8)] It is unclear whether the same text embedding y is used for every layer index i or whether each layer has its own text feature; please clarify the indexing and how the 40 adjectives are reduced to a scalar similarity for the regressor.
  5. [Figure 5] The reported FLOPS and parameter counts are only shown graphically; providing numerical values would make the complexity comparison quantitatively verifiable.
  6. [Abstract; Section I] The phrase 'achieves superior results' should be softened unless paired statistics are supplied, because the margins over the strongest baseline are within one reported standard deviation on several datasets.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: BPCLIP's reported results are supervised regressions to MOS on held-out splits, and the paper's components are not defined in terms of the outputs they predict.

full rationale

I walked the paper's derivation chain. The final score is computed in Eqs. (8)-(9) as a cosine similarity between image features and fixed CLIP text embeddings, followed by an MLP regression; the loss in Eq. (10) is a standard MSE against MOS. Nothing in this chain defines the text descriptors, the image features, or the attention modules in terms of the MOS target being predicted. The 40 quality adjectives are inputs to the frozen CLIP text encoder, not outputs derived from the benchmarks. No equation in the paper reduces a 'prediction' to the fitted data by construction, and no fitted parameter is renamed as a prediction. The cited prior work [11] supplies GLP and serves as the TOPIQ baseline, but BPCLIP's bottom-up MSCA and text-similarity scoring are implemented independently and evaluated against external benchmarks; there is no load-bearing self-citation chain. The paper does not invoke a uniqueness theorem or smuggle in an ansatz via a same-author citation. The main weaknesses—the unenumerated 40 adjectives, the lack of paired per-split baseline statistics, and the fact that gains over TOPIQ are often within one reported standard deviation—are reproducibility and statistical-significance concerns, not circularity. Supervised regression on held-out splits is a normal and non-circular evaluation protocol.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The central claim rests on supervised fitting of network weights to MOS scores and on a hand-designed, undisclosed text prompt set. No analytical derivation is offered; performance is an empirical result of fitted parameters.

free parameters (2)
  • Learnable weights in image encoder and score regressor = trained on MOS targets; numeric values not reported
    GLP bottleneck convolutions, MSCA and SA projections, learnable positional encoding, MLPs, and the final score MLP are all optimized with MSE loss on IQA datasets; reported SRCC/PLCC are results of this supervised fit.
  • 40 hand-selected image quality adjectives in six dimensions = not enumerated in the paper
    The adjective set is the text-side quality vocabulary; ablation Table VI shows it contributes to performance, but its exact contents and whether it was tuned on the benchmarks are not disclosed.
assumptions (4)
  • domain assumption MOS labels from benchmark datasets are valid scalar ground truth for perceived quality.
    All training and evaluation regress to normalized MOS in Eq. (10); if these labels are unreliable, the SRCC/PLCC claims inherit that unreliability.
  • ad hoc to paper Frozen CLIP text embeddings of the 40 quality adjectives remain semantically aligned with the fine-tuned image features.
    The paper relies on this for the cosine similarities in Eq. (8), but the prompts are not listed and only an ablation supports them.
  • domain assumption Bottom-up cross-scale attention propagates distortion information from low to high levels as intended.
    The qualitative claim is supported only by attention visualizations in Figure 4, with no analytical guarantee.
  • domain assumption Random 6:2:2 splits and reference-based cropping prevent content overlap between training and test sets.
    The protocol assumes no leakage; no analysis of duplicate or near-duplicate content is provided.
invented entities (1)
  • 40 image quality adjectives across six dimensions
    purpose: Fixed text prompts for the frozen CLIP text encoder to produce quality-related semantic features for cosine scoring.
    The list is not given, so it cannot be externally checked; independent evidence would require a public artifact or a prediction outside the fitted benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BPCLIP: A Bottom-up Image Quality Assessment from Distortion to Semantics Based on CLIP." pith.science (2026). https://pith.science/paper/L3PKZFCV

@misc{pith2026250617969,
  author       = {Pith},
  title        = {Pith review of: BPCLIP: A Bottom-up Image Quality Assessment from Distortion to Semantics Based on CLIP},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L3PKZFCV}},
  note         = {Machine review of arXiv:2506.17969}
}
read the original abstract

Image Quality Assessment (IQA) aims to evaluate the perceptual quality of images based on human subjective perception. Existing methods generally combine multiscale features to achieve high performance, but most rely on straightforward linear fusion of these features, which may not adequately capture the impact of distortions on semantic content. To address this, we propose a bottom-up image quality assessment approach based on the Contrastive Language-Image Pre-training (CLIP, a recently proposed model that aligns images and text in a shared feature space), named BPCLIP, which progressively extracts the impact of low-level distortions on high-level semantics. Specifically, we utilize an encoder to extract multiscale features from the input image and introduce a bottom-up multiscale cross attention module designed to capture the relationships between shallow and deep features. In addition, by incorporating 40 image quality adjectives across six distinct dimensions, we enable the pre-trained CLIP text encoder to generate representations of the intrinsic quality of the image, thereby strengthening the connection between image quality perception and human language. Our method achieves superior results on most public Full-Reference (FR) and No-Reference (NR) IQA benchmarks, while demonstrating greater robustness.

Figures

Figures reproduced from arXiv: 2506.17969 by the authors.

Figure 1
Figure 1. Overview of the final quality score computation strat [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architecture overview of the proposed BPCLIP Image Encoder. Five distinct levels of features were extracted from the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. GLP block comprises a mask branch and a feature [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Attention-based visualization of Gaussian blur distortion from the LIVE dataset. First row: the output [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Computational cost (FLOPS) vs. Performance (PLCC) [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CRISP-SAM2: SAM2 with Cross-Modal Interaction and Semantic Prompting for Multi-Organ Segmentation

    eess.IV 2025-06 conditional novelty 5.0 of 10

    A text-guided SAM2 variant with cross-modal attention, semantic prompt generation, and a similarity-sorted memory bank achieves top Dice and surface scores on seven public multi-organ CT datasets.

  2. Hierarchical Characterization of Brain Dynamics via State Space-based Vector Quantization

    eess.IV 2025-06 conditional novelty 5.0 of 10

    A hierarchical state-space vector-quantization model, HST, quantizes fMRI brain states and transitions into discrete tokens and reports modest classification gains on ADHD and schizophrenia datasets.

Reference graph

Works this paper leans on

39 extracted references · 36 canonical work pages · cited by 2 Pith papers

  1. [1]

    Learning generalizable perceptual representations for data-efficient no-reference image quality assessment,

    Suhas Srinath, Shankhanil Mitra, et al., “Learning generalizable perceptual representations for data-efficient no-reference image quality assessment,” inWACV, 2024, pp. 22–31

  2. [2]

    Exploring clip for assessing the look and feel of images,

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy, “Exploring clip for assessing the look and feel of images,” inAAAI, 2023, vol. 37, pp. 2555–2563

  3. [3]

    Blind image quality assessment via vision-language correspondence: A multitask learning perspective,

    Weixia Zhang, Guangtao Zhai, et al., “Blind image quality assessment via vision-language correspondence: A multitask learning perspective,” inCVPR, 2023, pp. 14071–14081

  4. [4]

    Learning transferable visual models from natural language supervision,

    Alec Radford, Jong Wook Kim, et al., “Learning transferable visual models from natural language supervision,” inICML. PMLR, 2021, pp. 8748–8763

  5. [5]

    Iterative prompt learning for unsupervised backlit image enhancement,

    Zhexin Liang, Chongyi Li, et al., “Iterative prompt learning for unsupervised backlit image enhancement,” inICCV, 2023, pp. 8094– 8103

  6. [6]

    Sglp: A similarity guided fast layer par- tition pruning for compressing large deep models,

    Yuqi Li, Yao Lu, et al., “Sglp: A similarity guided fast layer par- tition pruning for compressing large deep models,”arXiv preprint arXiv:2410.14720, 2024

  7. [7]

    Fedkd-hybrid: Federated hybrid knowl- edge distillation for lithography hotspot detection,

    Yuqi Li, Xingyou Lin, et al., “Fedkd-hybrid: Federated hybrid knowl- edge distillation for lithography hotspot detection,”arXiv preprint arXiv:2501.04066, 2025

  8. [8]

    Multiscale structural similarity for image quality assessment,

    Zhou Wang, Eero P Simoncelli, and Alan C Bovik, “Multiscale structural similarity for image quality assessment,” inACSSC. Ieee, 2003, vol. 2, pp. 1398–1402

Show all 39 references
  1. [9]

    No- reference image quality assessment in the spatial domain,

    Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik, “No- reference image quality assessment in the spatial domain,”IEEE TIP, vol. 21, no. 12, pp. 4695–4708, 2012

  2. [10]

    Making a “completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik, “Making a “completely blind” image quality analyzer,”IEEE SPL, vol. 20, no. 3, pp. 209–212, 2012

  3. [11]

    Topiq: A top-down approach from semantics to distortions for image quality assessment,

    Chaofeng Chen, Jiadi Mo, et al., “Topiq: A top-down approach from semantics to distortions for image quality assessment,”IEEE TIP, 2024

  4. [12]

    Deep residual learning for image recognition,

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” inCVPR, 2016, pp. 770–778

  5. [13]

    Attention is all you need in speech separation,

    Cem Subakan, Mirco Ravanelli, et al., “Attention is all you need in speech separation,” inICASSP. IEEE, 2021, pp. 21–25

  6. [14]

    Image quality assessment: from error visibility to structural similarity,

    Zhou Wang, Alan C Bovik, et al., “Image quality assessment: from error visibility to structural similarity,”IEEE TIP, vol. 13, no. 4, pp. 600–612, 2004

  7. [15]

    Image information and visual quality,

    Hamid R Sheikh and Alan C Bovik, “Image information and visual quality,”IEEE TIP, vol. 15, no. 2, pp. 430–444, 2006

  8. [16]

    Fsim: A feature similarity index for image quality assessment,

    Lin Zhang, Lei Zhang, et al., “Fsim: A feature similarity index for image quality assessment,”IEEE TIP, vol. 20, no. 8, pp. 2378–2386, 2011

  9. [17]

    Most apparent distortion: full- reference image quality assessment and the role of strategy,

    Eric C Larson and Damon M Chandler, “Most apparent distortion: full- reference image quality assessment and the role of strategy,”JEI, vol. 19, no. 1, pp. 011006–011006, 2010

  10. [18]

    Vsi: A visual saliency-induced index for perceptual image quality assessment,

    Lin Zhang, Ying Shen, and Hongyu Li, “Vsi: A visual saliency-induced index for perceptual image quality assessment,”IEEE TIP, vol. 23, no. 10, pp. 4270–4281, 2014

  11. [19]

    Deep learning of human visual sensitivity in image quality assessment framework,

    Jongyoo Kim and Sanghoon Lee, “Deep learning of human visual sensitivity in image quality assessment framework,” inCVPR, 2017, pp. 1676–1684

  12. [20]

    Pieapp: Perceptual image-error assessment through pairwise preference,

    Ekta Prashnani, Hong Cai, et al., “Pieapp: Perceptual image-error assessment through pairwise preference,” inCVPR, 2018, pp. 1808– 1817

  13. [21]

    The unreasonable effectiveness of deep features as a perceptual metric,

    Richard Zhang, Phillip Isola, et al., “The unreasonable effectiveness of deep features as a perceptual metric,” inCVPR, 2018, pp. 586–595

  14. [22]

    Image quality assessment: Unifying structure and texture similarity,

    Keyan Ding, Kede Ma, et al., “Image quality assessment: Unifying structure and texture similarity,”IEEE T-PAMI, vol. 44, no. 5, pp. 2567– 2581, 2020

  15. [23]

    Content-variant reference image quality assessment via knowledge distillation,

    Guanghao Yin, Wei Wang, et al., “Content-variant reference image quality assessment via knowledge distillation,” inAAAI, 2022, vol. 36, pp. 3134–3142

  16. [24]

    Maniqa: Multi-dimension attention network for no-reference image quality assessment,

    Sidi Yang, Tianhe Wu, et al., “Maniqa: Multi-dimension attention network for no-reference image quality assessment,” inCVPR, 2022, pp. 1191–1200

  17. [25]

    Incorporating semi-supervised and positive-unlabeled learning for boosting full reference image quality assessment,

    Yue Cao, Zhaolin Wan, et al., “Incorporating semi-supervised and positive-unlabeled learning for boosting full reference image quality assessment,” inCVPR, 2022, pp. 5851–5861

  18. [26]

    Attentions help cnns see better: Attention-based hybrid image quality assessment network,

    Shanshan Lao, Yuan Gong, et al., “Attentions help cnns see better: Attention-based hybrid image quality assessment network,” inCVPR, 2022, pp. 1140–1149

  19. [27]

    A statistical evaluation of recent full reference image quality assessment algorithms,

    Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik, “A statistical evaluation of recent full reference image quality assessment algorithms,” IEEE TIP, vol. 15, no. 11, pp. 3440–3451, 2006

  20. [28]

    Color image database tid2013: Peculiarities and preliminary results,

    Nikolay Ponomarenko, Oleg Ieremeiev, et al., “Color image database tid2013: Peculiarities and preliminary results,” inEUVIP. IEEE, 2013, pp. 106–111

  21. [29]

    Kadid-10k: A large-scale artificially distorted iqa database,

    Hanhe Lin, Vlad Hosu, and Dietmar Saupe, “Kadid-10k: A large-scale artificially distorted iqa database,” inQoMEX. IEEE, 2019, pp. 1–3

  22. [30]

    Massive online crowdsourced study of subjective and objective picture quality,

    Deepti Ghadiyaram and Alan C Bovik, “Massive online crowdsourced study of subjective and objective picture quality,”IEEE TIP, vol. 25, no. 1, pp. 372–387, 2015

  23. [31]

    Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment,

    Vlad Hosu, Hanhe Lin, et al., “Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment,”IEEE TIP, vol. 29, pp. 4041–4056, 2020

  24. [32]

    Perceptual quality assessment of smartphone photography,

    Yuming Fang, Hanwei Zhu, et al., “Perceptual quality assessment of smartphone photography,” inCVPR, 2020, pp. 3677–3686

  25. [33]

    Blind image quality assessment: From natural scene statistics to perceptual quality,

    Anush Krishna Moorthy and Alan Conrad Bovik, “Blind image quality assessment: From natural scene statistics to perceptual quality,”IEEE TIP, vol. 20, no. 12, pp. 3350–3364, 2011

  26. [34]

    A feature-enriched completely blind image quality evaluator,

    Lin Zhang, Lei Zhang, and Alan C Bovik, “A feature-enriched completely blind image quality evaluator,”IEEE TIP, vol. 24, no. 8, pp. 2579–2591, 2015

  27. [35]

    Image quality assessment using contrastive learning,

    Pavan C Madhusudana, Neil Birkbeck, et al., “Image quality assessment using contrastive learning,”IEEE TIP, vol. 31, pp. 4149–4161, 2022

  28. [36]

    Re-iqa: Unsuper- vised learning for image quality assessment in the wild,

    Avinab Saha, Sandeep Mishra, and Alan C Bovik, “Re-iqa: Unsuper- vised learning for image quality assessment in the wild,” inCVPR, 2023, pp. 5846–5855

  29. [37]

    Arniqa: Learning dis- tortion manifold for image quality assessment,

    Lorenzo Agnolucci, Leonardo Galteri, et al., “Arniqa: Learning dis- tortion manifold for image quality assessment,” inWACV, 2024, pp. 189–198

  30. [38]

    No reference opinion unaware quality assessment of authentically distorted images,

    Nithin C Babu, Vignesh Kannan, and Rajiv Soundararajan, “No reference opinion unaware quality assessment of authentically distorted images,” inWACV, 2023, pp. 2459–2468

  31. [39]

    Musiq: Multi-scale image quality transformer,

    Junjie Ke, Qifei Wang, et al., “Musiq: Multi-scale image quality transformer,” inICCV, 2021, pp. 5148–5157

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.