Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

EBA-AI: Ethics-Guided Bias-Aware AI for Efficient Underwater Image Enhancement and Coral Reef Monitoring

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read EBA-AI uses CLIP embeddings to decide which underwater images truly need enhancement, skipping clear frames and cutting reported GPU workload by 5-33% while keeping PSNR loss near 1 dB.

desk verdict A competent engineering combination with a load-bearing efficiency claim that is measured as a skip rate, not as actual GPU savings; the fairness and uncertainty parts are described but never evaluated. read the letter →

arxiv 2507.15036 v1 pith:OFSIXTYR submitted 2025-07-20 cs.CV cs.RO

Lyes Saad Saoud , Irfan Hussain This is my paper · ORCID
classification cs.CVcs.RO
keywords underwaterimageenhancementcoralreefmonitoringCLIPembeddingsdatasetbiasmitigationenergy-efficientAIadaptiveinferenceuncertaintyestimationexplainable
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

This paper proposes EBA-AI, a framework that adds a CLIP-based clarity gate in front of an underwater image enhancement network. The gate scores each incoming frame against semantic descriptions of water conditions and sends only low-quality frames through the enhancement model; clear frames are skipped. On the LSUI400, UIEB100, and Ocean_ex benchmarks the authors report that this selective processing cuts reported GPU workload by 18.75%, 33%, and 5% respectively, while the peak signal-to-noise ratio drops by roughly 1 dB or less. The paper also uses CLIP embedding entropy to detect dataset bias and reweight training samples, and it adds Monte Carlo dropout uncertainty estimates and Grad-CAM maps so operators can see why an enhancement was made. If the workload savings translate to real runtime and energy, the framework would make continuous, large-scale coral reef monitoring practical on limited computing budgets.

What carries the argument

The load-bearing mechanism is a CLIP similarity threshold used as a gating policy. CLIP, a vision-language model trained on image-text pairs, produces an embedding $F_k \leftarrow C(I_k)$ for each frame, and a confidence score $S_k \leftarrow \text{CLIP-Similarity}(F_k)$ says how closely the image resembles clear-water conditions; if $S_k > T$, the frame is skipped, otherwise it is sent through the enhancement model. The paper also proposes a degradation map $M(x,y) = |I(x,y)-I_{\text{local}}(x,y)|/(I_{\text{local}}(x,y)+\epsilon)$ and a dynamic depth function $d(x,y)=\min(D_{\max}, \alpha M(x,y)+\beta)$ for region-adaptive processing, plus a bias weight $\mathcal{L}_{\text{bias}}=\sum_i w_i \mathcal{L}_{\text{task}}(f(I_i),y_i)$ and MC-dropout variance $\sigma^2(I)$ for uncertainty. The argument equates the fraction of frames skipped with the percentage of GPU work saved.

What would settle it

Measure wall-clock time and GPU energy on the same GPU for the full LSUI400 test set, once with all frames enhanced and once with the CLIP gate active. If the actual time or energy saving is far below the reported 18.75% while the PSNR drop remains about 0.8 dB, the central efficiency claim fails; if the saving holds, the claim is supported.

Watch

Extended reading notes

Core claim

The central claim is that a CLIP-based selective enhancement policy can reduce the number of underwater images sent through a deep enhancement network by 5 to 33 percent with only a small loss in quality, and that this changes the feasibility of real-time coral reef monitoring. In the ablation on LSUI400, full processing reaches 27.20 dB PSNR and 0.879 structural similarity (SSIM), whereas CLIP-gated processing reaches 26.40 dB and 0.869 SSIM while skipping 18.75% of the workload; the paper restates this as a 3.89% PSNR drop for an 18.75% workload reduction. On UIEB100 the gate skips 33% of frames (22.62 to 21.99 dB), and on Ocean_ex it skips 5% (21.37 to 20.91 dB). The authors further argue that CLIP scoring is preferable to heuristic filters such as brightness thresholding because it captures semantic conditions like slight turbidity or low contrast, and they report that bias-aware reweighting improves generalization across environments.

Load-bearing premise

The load-bearing premise is that the percentage of skipped images equals the percentage of GPU work saved, meaning total runtime and energy scale linearly with the number of frames passed through the enhancement network; the paper gives no direct measurement of GPU utilization, wall-clock time, FLOPs, or energy use.

Editorial extensions

If this is right

  • On a mixed stream with one third clear frames, gating can skip roughly a third of the enhancement compute at a cost of about 0.6 dB PSNR on UIEB100.
  • The reported savings would compound with region-adaptive depth control: degraded frames could still be processed only at high-degradation regions.
  • The CLIP gate gives a semantic, not brightness-only, decision, so slightly turbid or low-contrast frames can still be routed to the enhancer.
  • MC-dropout variance and Grad-CAM attribution provide a per-frame confidence and a spatial reason for each enhancement, supporting human oversight in conservation decisions.

Reading between the lines

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

  • I infer that the real runtime gain depends on the enhancement network dominating the cost of CLIP scoring, which the paper does not measure.
  • A deployment-level corollary: the savings rate will track the clear-frame rate, so the method is most valuable in clear-water or mixed surveys and almost pointless in uniformly turbid or deep-sea surveys like Ocean_ex's 5%.
  • A testable extension would separate the two mechanisms, CLIP gating and change-guided depth control, and measure each one's wall-clock and energy contribution on the same hardware.
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

5 major / 6 minor

Summary. EBA-AI proposes a framework for underwater image enhancement that combines CLIP-based bias detection and reweighting, a selective 'change-guided' adaptive processing strategy, Monte Carlo Dropout uncertainty estimation, and Grad-CAM explainability. The method is evaluated on LSUI400, UIEB100, and Ocean_ex against seven baselines. The central quantitative claim is that CLIP-based selective enhancement reduces computational workload by 5-33% (reported as 'GPU savings') with only a small PSNR loss, and that this trade-off enables real-time marine monitoring. The paper also reports CLIP similarity scores and t-SNE visualizations to argue that existing datasets are biased toward clear-water conditions.

Significance. If the efficiency and quality trade-off were properly supported, a CLIP-based selective enhancement policy could be practically useful for bandwidth- and energy-constrained underwater monitoring, and the dataset-bias analysis addresses a real concern in coral-reef computer vision. The paper also has a transparent algorithmic description and covers several public benchmarks. However, the claimed contributions go well beyond what the experiments demonstrate: the headline computational savings are never directly measured, the bias-mitigation component is not evaluated, and the uncertainty/explainability components are not tested at all. The current evidence is therefore not sufficient to support the paper's central claims, though most gaps are in principle addressable with additional experiments.

major comments (5)
  1. [Section 4.2, Table 5] The 'GPU Savings %' column is not a measurement of GPU usage, wall-clock time, FLOPs, energy consumption, or frame rate. The values (18.75%, 33.00%, 5.00%) match the fraction of images skipped by the CLIP threshold in Algorithm 1, but the paper never reports that threshold T, the number of skipped images per dataset, or any direct runtime measurement. Because CLIP inference adds a fixed per-image cost, total workload cannot be assumed to scale linearly with the number of enhanced images; in the Ocean_ex case, where only 5% of images are skipped, the full pipeline is almost certainly more expensive than the baseline if CLIP overhead is included. The real-time feasibility claim therefore rests on an unmeasured quantity and an unstated linear-scaling assumption.
  2. [Section 4.2, Tables 3 and 4] The text claims that EBA-AI 'achieved an SSIM of 0.806 and a PSNR of 20.911, surpassing WaterNet and RAUNE-Net' on Ocean_ex, but Table 3 shows WaterNet at SSIM 0.843 / PSNR 21.744 and RAUNE-Net at SSIM 0.811 / PSNR 21.366, both higher than EBA-AI's values. Similarly, for UIEB100 the text claims 'the highest SSIM (0.821) and PSNR (21.988)', but Table 4 lists RAUNE-Net at SSIM 0.831 and PSNR 22.618. These contradictions undermine the claim that EBA-AI outperforms state-of-the-art methods and need to be corrected or the claims substantially weakened.
  3. [Section 3.1 and Section 4.1] The bias-mitigation component is never tested. Equations (1) and (2) define an entropy-based bias measure and a reweighted loss, and Section 4.1 reports CLIP similarity scores and a t-SNE plot, but there is no experiment showing that the reweighting changes model behavior, no before/after comparison on a fairness or performance metric, and no downstream classification or regression task that would demonstrate 'balanced representation' or 'fairness'. As written, the bias-aware contribution is an unvalidated proposal rather than a demonstrated component.
  4. [Section 3.2, Algorithm 1, and Table 5] The adaptive processing contribution includes the degradation map M(x,y) in Eq. (3) and the dynamic depth function d(x,y) in Eq. (4), but neither is present in Algorithm 1, and the ablation in Table 5 only varies a whole-image CLIP skip decision. There is no experiment isolating the effect of Eq. (3) or Eq. (4), no report of the coefficients alpha, beta, or the threshold T, and no ablation showing that the 'change-guided' or 'dynamic depth' mechanisms contribute to the reported savings. The efficiency results therefore cannot be attributed to the proposed adaptive processing scheme.
  5. [Section 3.3 and Section 4.2] Uncertainty estimation and explainability are described in Section 3.3 and mentioned in Section 4.2, but no experimental results are presented: there are no MC Dropout variance values, no calibration or reliability analysis of the uncertainty scores, no Grad-CAM visualizations, and no evaluation of whether explanations help human decision-making. The paper's claims about 'enhanced interpretability' and 'trust' are therefore unsupported by any evidence in the evaluation.
minor comments (6)
  1. [Abstract and Table 5] The abstract states that 'PSNR drops by a controlled 1.0 dB', but the PSNR drops in Table 5 are 0.80 dB (LSUI400), 0.63 dB (UIEB100), and 0.46 dB (Ocean_ex); please reconcile the stated drop with the reported numbers.
  2. [Eq. (1)] Equation (1) as written sums p(f(I_i)) log p(f(I_i)) over individual images; it is not a well-defined entropy unless p is specified as a distribution over embedding clusters or a non-uniform empirical distribution. Please clarify the definition.
  3. [Section 4] The training description says '100 iterations' but also mentions checkpoints every five epochs and validation at 500-iteration intervals; please state unambiguously whether training ran for 100 epochs or 100 iterations per epoch, and report the total training schedule.
  4. [Table 5] The 'Without CLIP' rows reproduce the RAUNE-Net rows from Tables 2-4 (e.g., LSUI400 27.20/0.879, UIEB100 22.62/0.831). Please state explicitly what the full-processing baseline is, namely whether the enhancement network is RAUNE-Net and whether 'Without CLIP' corresponds to EBA-AI with all other components but no filtering.
  5. [Section 4.2] The sentence reporting a 'minor 3.89% drop in PSNR' does not match Table 5, where the relative drops from the 'Without CLIP' baselines are approximately 2.9% (LSUI400), 2.8% (UIEB100), and 2.2% (Ocean_ex); please recompute or correct this percentage.
  6. [Section 6] The conclusion mentions 'improving classification fairness', but the paper does not perform a classification task; please rephrase to refer to enhancement-oriented fairness or remove the claim.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline GPU-savings claim reduces to the CLIP skip fraction by construction: 'GPU savings %' is the percentage of images skipped by an undisclosed threshold, not a measured computational saving.

  1. self definitional [Algorithm 1 (Section 3.2) and Section 4.2, Table 5]
    "if Sk > T then ▷ High confidence: Likely clear image ... Skip enhancement: S ← S ∪ {Ik} ... The GPU savings per dataset were as follows: LSUI400 (18.75%), UIEB100 (33.00%), and Ocean_ex (5.00%). These savings significantly reduced energy consumption, making the model viable for real-time applications in marine conservation."

    Algorithm 1 defines S as the set of images skipped by the CLIP confidence threshold T, so |S|/N is the fraction of images not processed by the enhancement network. The paper's central efficiency number, labeled 'GPU savings %' and then translated into 'reduces computational workload by 18.75%' and 'real-time feasibility', is exactly that same skip fraction. No GPU-utilization, wall-clock, FLOPs, or energy measurement is reported anywhere in Section 4. Since the threshold T is not disclosed and CLIP inference adds a per-image overhead, calling the skip rate 'GPU savings' makes the headline efficiency claim true by definition of the column rather than by measurement.

full rationale

The paper's quality evaluations (PSNR, SSIM, UIQM, etc.) are independent empirical measurements, and the bias/uncertainty discussion is mostly qualitative; those parts are not circular. The circularity is concentrated in the efficiency claim. Table 5 labels the fraction of images that the CLIP threshold chooses to skip as 'GPU Savings %', and the text converts that same fraction into a computational-workload reduction and real-time feasibility. Because the skipped fraction is generated by Algorithm 1's own threshold rule and is not validated against any runtime or energy measurement, the central '18.75% reduction' is definitional. The paper also does not account for the added cost of running CLIP on every image, so even the qualitative direction of the saving is not established. The 'Without CLIP' row in Table 5 reproduces the RAUNE-Net values from Tables 2-4, which introduces a baseline-identity ambiguity, though this is a comparison/mislabeling issue rather than circular reasoning. Self-citations [33-35] are present but not load-bearing. Overall, partial circularity: the efficiency headline is built into the table's column definition, while quality results retain independent content.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The framework rests on several undisclosed knobs and unvalidated proxies: the CLIP threshold T, dynamic-depth coefficients, and bias weights are never fixed; CLIP similarity is treated as ground truth for environmental bias; skipped-image fraction is treated as GPU savings; and the degradation map's benefit is asserted without ablation.

free parameters (4)
  • CLIP confidence threshold T = Not reported
    Algorithm 1 skips images with CLIP similarity above T; T controls the quality and efficiency tradeoff, so the reported savings and PSNR drops depend on an undisclosed hand-selected value.
  • Depth scaling coefficients alpha and beta = Not reported
    Eq. 4 defines dynamic depth d(x,y) = min(Dmax, alpha*M + beta), but no values, sweeps, or validation are given, and no experiment isolates this part of the framework.
  • Per-sample bias weights w_i = Not reported
    Eq. 2 defines contrastive reweighting via w_i, but the paper gives no formula for w_i and no evidence that reweighted training was performed; this is the core of the bias mitigation claim.
  • MC Dropout forward passes T = Not reported
    Eq. 5 uses T stochastic forward passes; no value or calibration is specified, and the uncertainty output is not evaluated.
assumptions (3)
  • domain assumption CLIP embedding similarity to text prompts reflects true environmental condition distribution in underwater images.
    Section 4.1 uses CLIP similarities and t-SNE to conclude which datasets are biased; this presumes CLIP's zero-shot similarity scores are calibrated measures of water clarity and depth, which is not validated.
  • domain assumption GPU workload scales linearly with the number of processed images, so skipped-image fraction equals energy and GPU savings.
    Table 5 and the GPU savings paragraph convert the percentage of images skipped by CLIP into 'GPU savings %' without measuring utilization or energy.
  • ad hoc to paper Local contrast difference M(x,y) in Eq. 3 identifies regions that require enhancement and supports the dynamic depth policy.
    Section 3.2 introduces Eqs. 3 and 4, but no experiment isolates the degradation map or verifies that lightweight processing of low-degradation regions preserves quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EBA-AI: Ethics-Guided Bias-Aware AI for Efficient Underwater Image Enhancement and Coral Reef Monitoring." pith.science (2026). https://pith.science/paper/OFSIXTYR

@misc{pith2026250715036,
  author       = {Pith},
  title        = {Pith review of: EBA-AI: Ethics-Guided Bias-Aware AI for Efficient Underwater Image Enhancement and Coral Reef Monitoring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OFSIXTYR}},
  note         = {Machine review of arXiv:2507.15036}
}
read the original abstract

Underwater image enhancement is vital for marine conservation, particularly coral reef monitoring. However, AI-based enhancement models often face dataset bias, high computational costs, and lack of transparency, leading to potential misinterpretations. This paper introduces EBA-AI, an ethics-guided bias-aware AI framework to address these challenges. EBA-AI leverages CLIP embeddings to detect and mitigate dataset bias, ensuring balanced representation across varied underwater environments. It also integrates adaptive processing to optimize energy efficiency, significantly reducing GPU usage while maintaining competitive enhancement quality. Experiments on LSUI400, Oceanex, and UIEB100 show that while PSNR drops by a controlled 1.0 dB, computational savings enable real-time feasibility for large-scale marine monitoring. Additionally, uncertainty estimation and explainability techniques enhance trust in AI-driven environmental decisions. Comparisons with CycleGAN, FunIEGAN, RAUNENet, WaterNet, UGAN, PUGAN, and UTUIE validate EBA-AI's effectiveness in balancing efficiency, fairness, and interpretability in underwater image processing. By addressing key limitations of AI-driven enhancement, this work contributes to sustainable, bias-aware, and computationally efficient marine conservation efforts. For interactive visualizations, animations, source code, and access to the preprint, visit: https://lyessaadsaoud.github.io/EBA-AI/

Figures

Figures reproduced from arXiv: 2507.15036 by the authors.

Figure 1
Figure 1. Illustration of the EBA-AI framework. Input images are filtered by CLIP [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualization of dataset diversity based on CLIP embeddings. Each [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Comparison of image enhancement results across multiple models for the [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of image enhancement results across multiple models for the [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Comparison of image enhancement results across multiple models for the [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. FishDet-M: A Unified Large-Scale Benchmark for Robust Fish Detection and CLIP-Guided Model Selection in Diverse Aquatic Visual Domains

    cs.CV 2025-07 conditional novelty 5.0 of 10

    FishDet-M merges 13 fish detection datasets into a 105,556-image COCO-style benchmark and benchmarks 28 detectors, with YOLO12x reaching the highest mAP of 0.491.

Reference graph

Works this paper leans on

45 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Discover Artificial Intelligence (2024)

    Alotaibi, E., Nassif, N.: Artificial intelligence in environmental monitoring: in- depth analysis. Discover Artificial Intelligence (2024)

    work page 2024

  2. [2]

    In: ICASSP - IEEE International Conference on Acoustics, Speech and Signal Processing (2024)

    Bakhtiarifard, P., Igel, C., Selvan, R.: Ec-nas: Energy consumption aware tabu- lar benchmarks for neural architecture search. In: ICASSP - IEEE International Conference on Acoustics, Speech and Signal Processing (2024)

    work page 2024

  3. [3]

    IEEE Access (2024)

    Barbierato, E., Gatti, A.: Toward green ai: A methodological survey of the scientific literature. IEEE Access (2024)

    work page 2024

  4. [4]

    International Journal of Re- mote Sensing (2017)

    Chegoonian, A., Mokhtarzade, M., Valadan Zoej, M.: A comprehensive evaluation of classification algorithms for coral reef habitat mapping: Challenges related to quantity, quality, and impurity of training samples. International Journal of Re- mote Sensing (2017)

    work page 2017

  5. [5]

    IEEE Transactions on Image Processing 32, 4472–4485 (2023)

    Cong, R., Yang, W., Zhang, W., Li, C., Guo, C.L., Huang, Q., Kwong, S.: Pu- gan: Physical model-guided underwater image enhancement using gan with dual- discriminators. IEEE Transactions on Image Processing 32, 4472–4485 (2023). https://doi.org/10.1109/TIP.2023.3286263

    work page arXiv 2023

  6. [6]

    Lecture Notes in Computer Science (2024)

    Cortês, G., Lourenço, N., Machado, P.: Towards physical plausibility in neuroevo- lution systems. Lecture Notes in Computer Science (2024)

    work page 2024

  7. [7]

    Sustainable Computing: Informatics and Systems (2023)

    Desislavov, R., Martínez-Plumed, F., Hernández-Orallo, J.: Trends in ai inference energy consumption: Beyond the performance-vs-parameter laws of deep learning. Sustainable Computing: Informatics and Systems (2023)

    work page 2023

  8. [8]

    Frontiers in Marine Science (2022)

    Ditria, E., Buelow, C., Gonzalez-Rivero, M., Connolly, R.: Artificial intelligence and automated monitoring for assisting conservation of marine ecosystems: A per- spective. Frontiers in Marine Science (2022)

    work page 2022

Show all 45 references
  1. [9]

    Communications in Computer and Information Science (2024)

    Duong, M., Conrad, S.: Towards fairness and privacy: A novel data pre-processing optimization framework for non-binary protected attributes. Communications in Computer and Information Science (2024)

    2024

  2. [10]

    In: 2018 IEEE International Con- ference on Robotics and Automation (ICRA)

    Fabbri, C., Islam, M.J., Sattar, J.: Enhancing underwater imagery us- ing generative adversarial networks. In: 2018 IEEE International Con- ference on Robotics and Automation (ICRA). pp. 7159–7165 (2018). https://doi.org/10.1109/ICRA.2018.8460552

    2018

  3. [11]

    NPJ Ocean Sustain- ability (2023)

    González-Espinosa, P., Bossier, S., Singh, G., Cisneros-Montemayor, A.: Integrat- ing equity-focused planning into coral bleaching management. NPJ Ocean Sustain- ability (2023)

    2023

  4. [12]

    In: Exploring Ethical Dimensions of En- vironmental Sustainability and Use of AI (2023)

    Gundeti, R., Vuppala, K., Kasireddy, V.: The future of ai and environmental sus- tainability: Challenges and opportunities. In: Exploring Ethical Dimensions of En- vironmental Sustainability and Use of AI (2023)

    2023

  5. [13]

    Coral Reefs (2024)

    Hoadley, K., Lowry, S., McQuagge, A., Muller, E.: Bio-optical signatures of in situ photosymbionts predict bleaching severity prior to thermal stress in the caribbean coral species acropora palmata. Coral Reefs (2024)

    2024

  6. [14]

    In: Lecture Notes in Networks and Systems (2024) 14 Lyes Saad Saoud and Irfan Hussain

    Iftikhar, S., Davy, S.: Reducing carbon footprint in ai: A framework for sustainable training of large language models. In: Lecture Notes in Networks and Systems (2024) 14 Lyes Saad Saoud and Irfan Hussain

    2024

  7. [15]

    IEEE Robotics and Automation Letters5(2), 3227–3234 (2020)

    Islam, M.J., Xia, Y., Sattar, J.: Fast underwater image enhancement for improved visual perception. IEEE Robotics and Automation Letters5(2), 3227–3234 (2020)

    2020

  8. [16]

    Journal of Applied Ecology (2023)

    Jackett, C., Althaus, F., Maguire, K., Williams, A.: A benthic substrate classifi- cation method for seabed images using deep learning: Application to management of deep-sea coral reefs. Journal of Applied Ecology (2023)

    2023

  9. [17]

    Conservation Science Western Australia (2012)

    Josephitis, E., Wilson, S., Moore, J., Field, S.: Comparison of three digital im- age analysis techniques for assessment of coral cover and bleaching. Conservation Science Western Australia (2012)

    2012

  10. [18]

    Neural Computing and Applications (2024)

    Karamchandani, A., Mozo, A., Gómez-Canaval, S., Pastor, A.: A methodological framework for optimizing the energy consumption of deep neural networks: a case study of a cyber threat detector. Neural Computing and Applications (2024)

    2024

  11. [19]

    In: IDIMT 2024: Changes to ICT, Management, and Busi- ness Processes through AI - 32nd Interdisciplinary Information Management Talks (2024)

    Kuchtíková, N., Maryska, M.: Eco-friendly ai: Balancing innovation with environ- mental responsibility. In: IDIMT 2024: Changes to ICT, Management, and Busi- ness Processes through AI - 32nd Interdisciplinary Information Management Talks (2024)

    2024

  12. [20]

    Proceedings of the Royal Society B: Biological Sciences (2018)

    Lam, V., Chaloupka, M., Thompson, A., Mumby, P.: Acute drivers influence recent inshore great barrier reef dynamics. Proceedings of the Royal Society B: Biological Sciences (2018)

    2018

  13. [21]

    IEEE Transactions on Image Processing 29, 4376–4389 (2019)

    Li, C., Guo, C., Ren, W., Cong, R., Hou, J., Kwong, S., Tao, D.: An underwater image enhancement benchmark dataset and beyond. IEEE Transactions on Image Processing 29, 4376–4389 (2019)

    2019

  14. [22]

    Science of the Total Environment (2025)

    Liu, L., Bao, Z., Liang, Y., Zhang, Z.: Unsupervised learning for lake underwater vegetation classification: Constructing high-precision, large-scale aquatic ecological datasets. Science of the Total Environment (2025)

    2025

  15. [23]

    In: INDICON 2022 - 2022 IEEE 19th India Council International Conference (2022)

    Maan, J.: Deep learning-driven explainable ai using generative adversarial network (gan). In: INDICON 2022 - 2022 IEEE 19th India Council International Conference (2022)

    2022

  16. [24]

    Ecosphere (2024)

    McClanahan, T., Azali, M., Muthiga, N., Guillaume, M.: Complex multivariate model predictions for coral diversity with climatic change. Ecosphere (2024)

    2024

  17. [25]

    In: Oceans Conference Record (IEEE) (2024)

    Minai, G., Bobeldyk, D., Leidig, J.: Evaluating the impact of diverse marine en- vironments on image classification performance. In: Oceans Conference Record (IEEE) (2024)

    2024

  18. [26]

    Water (Switzerland) (2022)

    Mukonza, S., Chiang, J.L.: Micro-climate computed machine and deep learning models for prediction of surface water temperature using satellite data in mundan water reservoir. Water (Switzerland) (2022)

    2022

  19. [27]

    International Journal on Informatics Visualization (2024)

    Nguyen, H., Nguyen, C., Tran, T., Pham, N.: Artificial intelligence and machine learning for green shipping: Navigating towards sustainable maritime practices. International Journal on Informatics Visualization (2024)

    2024

  20. [28]

    In: Proceedings of SPIE - The International Society for Optical Engineering (2019)

    Nurdin, N., Lanuru, M., Jalil, A., Komatsu, T.: Integration in-situ measurement and medium resolution imagery to develop digital health chart: Preliminary study of coral reefs on small islands, spermonde archipelago, indonesia. In: Proceedings of SPIE - The International Socie...

    2019

  21. [29]

    Page, C., Field, S., Pollock, F., Wilson, S.: Assessing coral health and disease from digitalphotographsandinsitusurveys.EnvironmentalMonitoringandAssessment (2017)

    2017

  22. [30]

    In: ACM International Conference Proceeding Series (2022)

    Pastaltzidis, I., Dimitriou, N., Quezada-Tavarez, K., Tzovaras, D.: Data augmen- tation for fairness-aware machine learning: Preventing algorithmic bias in law en- forcement systems. In: ACM International Conference Proceeding Series (2022)

    2022

  23. [31]

    IEEE Transactions on Image Processing 32, 3066–3079 (2023)

    Peng, L., Zhu, C., Bian, L.: U-shape transformer for underwater image en- hancement. IEEE Transactions on Image Processing 32, 3066–3079 (2023). https://doi.org/10.1109/TIP.2023.3276332 Title Suppressed Due to Excessive Length 15

    2023

  24. [32]

    In: Zhai, G., Zhou, J., Ye, L., Yang, H., An, P., Yang, X

    Peng, W., Zhou, C., Hu, R., Cao, J., Liu, Y.: Raune-net: A residual and attention- driven underwater image enhancement method. In: Zhai, G., Zhou, J., Ye, L., Yang, H., An, P., Yang, X. (eds.) Digital Multimedia Communications. pp. 15–27. Springer Nature Singapore, Singapore (2024)

    2024

  25. [33]

    IEEE Access 12, 145206–145233 (2024)

    Saad Saoud, L., Elmezain, M., Sultan, A., Heshmat, M., Seneviratne, L., Hussain, I.: Seeing through the haze: A comprehensive review of underwa- ter image enhancement techniques. IEEE Access 12, 145206–145233 (2024). https://doi.org/10.1109/ACCESS.2024.3465550

    2024

  26. [34]

    In: 2024 IEEE International Conference on Image Processing (ICIP)

    Saad Saoud, L., Niu, Z., Seneviratne, L., Hussain, I.: Real-time and resource-efficient multi-scale adaptive robotics vision for underwater ob- ject detection and domain generalization. In: 2024 IEEE International Conference on Image Processing (ICIP). pp. 3917–3923 (Oct 2024)...

    2024

  27. [35]

    In: 2023 21st International Conference on Advanced Robotics (ICAR)

    Saad Saoud, L., Niu, Z., Sultan, A., Seneviratne, L., Hussain, I.: Adod: Adaptive domain-aware object detection with residual attention for underwater environ- ments. In: 2023 21st International Conference on Advanced Robotics (ICAR). pp. 633–638 (Dec 2023). https://doi.org/10...

    2023

  28. [36]

    Lecture Notes in Computer Science (2023)

    Selvan, R., Schön, J., Dam, E.: Operating critical machine learning models in resource constrained regimes. Lecture Notes in Computer Science (2023)

    2023

  29. [37]

    Ecological Informatics (2025)

    Suan, A., Franceschini, S., Madin, J., Madin, E.: Quantifying 3d coral reef struc- tural complexity from 2d drone imagery using artificial intelligence. Ecological Informatics (2025)

    2025

  30. [38]

    Chemical Engineering Transactions (2023)

    Szarmes,P.,Élő,G.:Sustainabilityoflargeaimodels:Balancingenvironmentaland social impact with technology and regulations. Chemical Engineering Transactions (2023)

    2023

  31. [39]

    PeerJ (2019)

    Vallès, H., Oxenford, H., Henderson, A.: Switching between standard coral reef benthic monitoring protocols is complicated: Proof of concept. PeerJ (2019)

    2019

  32. [40]

    Sensors (2022)

    Villon, S., Iovan, C., Mangeas, M., Vigliola, L.: Confronting deep-learning and bio- diversity challenges for automatic video-monitoring of marine ecosystems. Sensors (2022)

    2022

  33. [41]

    PLoS ONE (2017)

    Welle, P., Small, M., Doney, S., Azevedo, I.: Estimating the effect of multiple environmental stressors on coral bleaching and mortality. PLoS ONE (2017)

    2017

  34. [42]

    Frontiers in Marine Science (2019)

    Williams, I., Couch, C., Beijbom, O., Brainard, R.: Leveraging automated image analysis tools to transform our capacity to assess status and trends on coral reefs. Frontiers in Marine Science (2019)

    2019

  35. [43]

    Xu, Y., Martínez-Fernández, S., Martinez, M., Franch, X.: Energy effi- ciency of training neural network architectures: An empirical study (2023), https://arxiv.org/abs/2302.00967

    2023 arXiv

  36. [44]

    Scientia Sinica Informationis (2024)

    Zheng, Z., Lü, J., Wang, L., Wei, W.: Cross-scale systematic learning for social big data: theory and methods. Scientia Sinica Informationis (2024)

    2024

  37. [45]

    In: Proceedings of the IEEE interna- tional conference on computer vision

    Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE interna- tional conference on computer vision. pp. 2223–2232 (2017)

    2017

Pith tools

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