Pith. sign in

REVIEW 5 major objections 5 minor 48 references

Improve Underwater Object Detection through YOLOv12 Architecture and Physics-informed Augmentation

T0 review · 5 major / 5 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 The paper claims that YOLOv12, augmented with physics-informed transformations, achieves state-of-the-art underwater object detection—98.30% mAP@0.5 on Brackish at 142 FPS—with gains of up to 10.46 percentage points over prior detectors.

desk verdict A plausible application paper whose headline efficiency claim is contradicted by its own speed table; the mAP results may be real, but the 142 FPS story needs rewriting. read the letter →

arxiv 2506.23505 v1 pith:Z7J6BF7Q submitted 2025-06-30 cs.CV

Tinh Nguyen This is my paper
classification cs.CV
keywords underwaterobjectdetectionYOLOv12physics-informedaugmentationAreaAttentionR-ELANreal-timeBrackishdatasetablationstudy
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 tries to establish that a real-time object detector, YOLOv12, combined with augmentations that simulate underwater light physics, can outperform existing underwater detectors on accuracy while keeping real-time speed. On four benchmark datasets (UDD, Brackish, URPC2019, DUO), the authors report mAP@0.5 values of 69.44%, 98.30%, 87.16%, and 87.44% respectively, all at 142 FPS. The key claim is that the attention-centric YOLOv12 architecture preserves structure in turbid water, and that the physics-informed augmentation recipe—spectral HSV shifts, depth-dependent blur, biological occlusion erasing, and symmetry-constrained flips—is what closes the gap to prior work. If true, this gives underwater robotics a drop-in detector that does not need a separate image-enhancement stage.

What carries the argument

The central mechanism is YOLOv12's attention-centric architecture paired with a four-step physics-informed augmentation pipeline. R-ELAN (Residual Efficient Layer Aggregation) blocks route features through parallel streams with varying numbers of Area Attention blocks, then add the original features back through a residual connection; this preserves structural information and stabilizes gradient flow in turbid images. Area Attention (A2) partitions the feature map into non-overlapping areas, reducing self-attention complexity from quadratic to $O((H\times W)^2/l)$ while keeping a receptive field large enough to infer occluded shapes. The augmentation pipeline applies HSV hue/saturation/value shifts modeled on spectral absorption, controlled blurring with depth-dependent kernels, random erasing that mimics power-law-distributed biological occlusion, and horizontal flips constrained to preserve bilateral symmetry. These components together simulate wavelength-dependent attenuation, scattering, and occlusion so the model learns robust shape inference from degraded inputs.

What would settle it

Retrain the leading baselines (SCR-Net, YOLOv9s-SD, EPBC-YOLOv8, Vanilla-YOLO) on the same data splits with the same 640×640 input, same augmentation, and same training schedule; if the mAP@0.5 gaps shrink to near zero or reverse, the SOTA claim is falsified. A simpler check runs the released YOLOv12 model on a single edge GPU to see if 142 FPS is reproducible outside the two-A100 setup.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the YOLOv12 architecture transfers to the underwater domain when its training data is perturbed in ways that mirror water's optical physics. The authors report state-of-the-art mAP@0.5 on all four evaluated datasets, with the largest margins on URPC2019 (+10.46 points over EPBC-YOLOv8) and UDD (+7.94 points over SCR-Net). The ablation on Brackish shows a 9.0-point jump in mAP@0.5 (89.3% to 98.30%) from adding the four augmentations in sequence, which the authors read as evidence that the augmentation strategy is the crucial ingredient. They also claim specific robustness benefits: 18.9% higher precision on occluded objects, 22.4% higher recall on sub-50-pixel objects, and a 34% inference speed-up from the FlashAttention integration.

Load-bearing premise

The state-of-the-art conclusion rests on comparing the authors' results with mAP numbers quoted from other papers, which used different training protocols, input sizes, and hardware; if those baselines are not comparable, the claimed improvements of 7.94–10.46 percentage points are not supported.

Editorial extensions

If this is right

  • On the paper's numbers, a single model reaches state-of-the-art mAP@0.5 on UDD, Brackish, URPC2019, and DUO while running at 142 FPS, so the accuracy-efficiency trade-off that motivated the work is resolved for these benchmarks.
  • The ablation attributes an increase from 89.3% to 98.30% mAP@0.5 to the sequential addition of horizontal flip, controlled blurring, HSV transformations, and random erasing, making the augmentation strategy the paper's main claimed contribution.
  • The claimed speed advantage (142 FPS versus 89–117 FPS for most baselines) suggests YOLOv12-UnderWater is deployable for real-time AUV tasks, provided the hardware is comparable to the two-A100 setup used in training.
  • The paper's own mAP@0.5:0.95 numbers trail some baselines (74.14% vs 77.60% on Brackish; 56.34% vs 64.00% on DUO), so the reported lead is at coarse IoU thresholds rather than precise localization.

Reading between the lines

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

  • The paper does not isolate the architecture's contribution from the augmentation's contribution beyond one ablation on Brackish; a controlled experiment that applies the same augmentation recipe to YOLOv8 or YOLOv11 would reveal how much of the 9-point gain is architecture-specific.
  • Because the augmentation parameters (e.g., $\sigma(z)\propto z^{0.78}$) are tuned to homogeneous water conditions, a natural extension is to make them spatially or depth-varying; the paper explicitly flags extreme turbidity gradients as out of scope.
  • The 142 FPS figure, cited as running on two A100 GPUs, has not been verified on AUV-embedded processors; a useful next step is a latency benchmark on a single edge GPU, which would test the deployment claim.
  • The reported 22.4% small-object recall and 18.9% occluded-object precision gains, if reproducible, would make the detector a promising front-end for automated marine biodiversity surveys, though the paper does not evaluate that use case.
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 / 5 minor

Summary. The paper proposes an underwater object detection pipeline built on the YOLOv12 architecture, augmented by physics-inspired image transformations (HSV spectral adjustments, controlled blurring, random erasing, and horizontally constrained flips), and evaluates it on four public underwater benchmarks (UDD, Brackish, URPC2019, DUO). The headline claims state-of-the-art results, including 98.30% mAP@0.5 on Brackish at 142 FPS, plus improved occlusion robustness (18.9%) and small-object recall (22.4%) over prior methods. The manuscript includes ablation studies of the augmentation strategy, loss weighting, optimizer choice, and input resolution.

Significance. If the results hold, the paper would offer a strong practical result: a real-time attention-based detector with a simple, domain-motivated augmentation recipe that outperforms many existing underwater detectors on standard benchmarks. The authors are also transparent about some limitations (homogeneous-water assumption, unverified edge-device performance, benthic-only evaluation). However, the current evidence does not yet support the central claims because of internal contradictions in the efficiency numbers, non-comparable baselines, and missing quantitative support for the headline percentage improvements. The work is potentially useful, but it needs substantial additional validation and reporting before its conclusions can be accepted.

major comments (5)
  1. [Tables 2–5; §4.3] The central efficiency claim is internally inconsistent. The abstract and the comparison tables (Tables 2–5) report YOLOv12 at 142 FPS, while Table 9 reports 610 FPS at the 640×640 input used throughout the experiments, 150 FPS at 1280×1280 (with mAP@0.5 dropping to 96.6%), and 1100 FPS at 320×320. No row of Table 9 produces 98.30% mAP@0.5 at 142 FPS. Moreover, §3.1 attributes 142 FPS to "embedded AUV hardware," §4.1 states that all experiments used two NVIDIA A100 GPUs, and §5 concedes that 142 FPS "requires powerful GPUs" and that edge-device performance "has not been verified." These statements cannot all be correct. Because the claimed advantage over YOLOv8-MU in §4.3 (URPC2019) is explicitly framed as a 34% higher framerate (105 vs 142 FPS), this inconsistency is load-bearing for the accuracy-efficiency trade-off, not merely a presentational typo.
  2. [Tables 2–5 and §4.3] The state-of-the-art comparisons are based on accuracy and speed numbers taken from different original papers, with no common evaluation protocol. The baselines were presumably trained and tested on different hardware, input resolutions, batch sizes, epochs, augmentations, and possibly different train/test splits. For example, Table 4 cites YOLOv8-MU at 88.10 mAP and 105 FPS from another publication, while Table 5 cites Vanilla-YOLO at 86.60 mAP and 155 FPS. Without re-running these methods under the same protocol (or at least a clearly stated controlled comparison), the claimed improvements (7.94, 10.46, 5.65, 0.84 percentage points, etc.) are not established. This is the direct basis for the paper's "state-of-the-art" conclusion, so it requires either a unified re-evaluation or a different, appropriately qualified claim.
  3. [Abstract, §3.1, §4.3] The headline improvements of 18.9% in occlusion robustness and 22.4% in small-object recall appear only in prose and are not backed by any table, figure, or defined metric. The paper does not state how occlusion robustness or small-object recall was measured, on which exact subsets of which datasets, or with which baselines the comparison is made. In addition, the 18.9% figure is attributed to the AUDD dataset in §3.1, but AUDD is not among the four datasets listed in §4.2 (UDD, Brackish, URPC2019, DUO) and does not appear in any experimental table. The authors should provide corresponding per-category or sub-group results in a table with clear definitions, or remove these quantitative claims from the abstract.
  4. [§4.4, Tables 6–8] The ablation study is performed only on the Brackish dataset, and it does not separate the contribution of the YOLOv12 architecture from the contribution of the physics-informed augmentations. Table 6 starts from "No Augmentation" at 89.3% mAP, but it is not reported whether this baseline is YOLOv12 without augmentations or a different architecture. Consequently, the paper does not demonstrate that the proposed augmentations are what makes YOLOv12 outperform previous underwater detectors; it only shows that, for one dataset and one architecture, adding these augmentations in sequence improves a single mAP number. A two-factor ablation (architecture × augmentation) across at least two datasets is needed to support the paper's central attribution of gains to the combination.
  5. [Tables 2–8] All reported metrics are single runs with no error bars, no repeated trials, and no statistical significance tests. This matters because some of the claimed differences are small: for example, Table 5 reports 0.84 percentage point advantage over Vanilla-YOLO, and Table 7 reports differences of 0.3–0.8 percentage points between loss weighting schemes. Without multiple seeds or confidence intervals, these conclusions are indistinguishable from noise, especially on underwater data with significant class imbalance and annotation noise. The manuscript should at least report mean and standard deviation over several training runs for the main comparison and ablation tables.
minor comments (5)
  1. [§3.2, Eq. (5)] Equation (5) uses the second partial derivative of an object O, but O is never defined, and the variable τ is said to adapt to turbulence without a precise formula. Please provide the definition and explain how the resulting Perase map is applied to the image.
  2. [§3.3] The text says "Brackish, AUDD, and UPPC2019," but the datasets used later are UDD, Brackish, URPC2019, and DUO. "UPPC2019" appears to be a typo for URPC2019, and AUDD should either be removed or defined.
  3. [Table 2] The row for "YOLOv8n baseline [37]" lists FPS as 98.30, which is numerically identical to the headline mAP value 98.30 on Brackish; this is confusing. Please use consistent formatting and avoid a number that can be misread as a percentage.
  4. [§4.3] The phrase "with only 610 GFLOPS computational cost" is ambiguous: GFLOPs is typically a per-inference compute measure, not a cost unit comparable to FPS. Please clarify whether this is the model's FLOP count and how it relates to the FPS numbers.
  5. [§3.1, last paragraph] The sentence "accelerating inference by 34% compared to standard self-attention" is stated as a fact without a citation or a measurement in the paper. If this is a known property of FlashAttention, cite the relevant benchmark; otherwise report the measurement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the reported mAP/FPS figures are empirical benchmark evaluations, not consequences of the model or augmentation definitions, and the paper's inconsistencies are correctness issues, not circularity.

full rationale

I walked the claimed derivation chain from the methodology to the results. The YOLOv12 components (R-ELAN, Area Attention) are taken from the external YOLOv12 preprint [10], described by equations (1)-(4), and no load-bearing claim is justified by self-citation. The physics-informed augmentations are defined by explicit physical models (spectral hue-shift integral, Beer-Lambert value decay, PSF convolution, structured erasure), and their contribution is tested by the Table 6 ablation, which starts from a no-augmentation baseline and adds components sequentially. The loss-weighting and optimizer comparisons in Tables 7-8 are standard model selection, not predictions forced by construction. The headline '98.30% mAP at 142 FPS' is an empirical result, not an output derived from the paper's equations or fitted parameters. The substantive weaknesses are internal consistency (Table 9 reports 610 FPS at 640x640 and 150 FPS at 1280x1280, so no row yields 142 FPS; Section 5 concedes that 'the 142 FPS benchmark requires powerful GPUs; the performance of edge devices ... has not been verified') and reliance on externally reported baseline numbers for the state-of-the-art claim. These are reproducibility/comparability concerns, not circularity. I therefore find no circular step and score 0.

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

The paper introduces no new physical entities. It relies on hand-chosen hyperparameters (loss weights, resolution, augmentation constants) and on an unverified correspondence between symbolic physics formulas and the actual augmentation pipeline. It also assumes that numbers from other papers are directly comparable, which is a common but unvalidated assumption in benchmark papers.

free parameters (4)
  • Loss weighting ratios (lambda_cls:lambda_obj:lambda_box) = 7.5:0.5:1.5
    Selected via ablation to maximize mAP@0.5 on Brackish; reported as the best configuration without error bars or statistical testing.
  • Depth attenuation exponent = 0.78 in sigma(z) proportional to z^0.78
    Introduced in controlled blurring augmentation to match empirical attenuation; no source or independent measurement provided.
  • Input resolution = 640x640
    Chosen from resolution ablation as best speed-accuracy trade-off; the same resolution is used for all reported benchmark numbers.
  • Augmentation hyperparameters = Not fully specified (HSV ranges, blur kernel sizes, erasing probabilities)
    The paper provides symbolic formulas but no exact implementation values, so the actual augmentation parameters are undisclosed free choices.
assumptions (3)
  • domain assumption Baseline numbers from prior papers are directly comparable despite different training protocols and hardware.
    Section 4.3 compares against numbers reported in other papers that likely use different training setups, input sizes, and hardware, without a common evaluation protocol.
  • ad hoc to paper The physics formulas in Section 3.2 correspond to the actual augmentation implementation.
    Symbolic formulas such as Eq. 7's PSF are not connected to Eq. 8's Gaussian kernel, and no implementation details are provided to confirm the mapping.
  • domain assumption Dataset splits and annotation quality are consistent across compared methods.
    URPC2019 test annotations are unavailable; the paper says a custom partition is required but does not specify it, so comparisons on that dataset rely on an unknown split.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improve Underwater Object Detection through YOLOv12 Architecture and Physics-informed Augmentation." pith.science (2026). https://pith.science/paper/Z7J6BF7Q

@misc{pith2026250623505,
  author       = {Pith},
  title        = {Pith review of: Improve Underwater Object Detection through YOLOv12 Architecture and Physics-informed Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z7J6BF7Q}},
  note         = {Machine review of arXiv:2506.23505}
}
read the original abstract

Underwater object detection is crucial for autonomous navigation, environmental monitoring, and marine exploration, but it is severely hampered by light attenuation, turbidity, and occlusion. Current methods balance accuracy and computational efficiency, but they have trouble deploying in real-time under low visibility conditions. Through the integration of physics-informed augmentation techniques with the YOLOv12 architecture, this study advances underwater detection. With Residual ELAN blocks to preserve structural features in turbid waters and Area Attention to maintain large receptive fields for occluded objects while reducing computational complexity. Underwater optical properties are addressed by domain-specific augmentations such as turbulence adaptive blurring, biologically grounded occlusion simulation, and spectral HSV transformations for color distortion. Extensive tests on four difficult datasets show state-of-the-art performance, with Brackish data registering 98.30% mAP at 142 FPS. YOLOv12 improves occlusion robustness by 18.9%, small-object recall by 22.4%, and detection precision by up to 7.94% compared to previous models. The crucial role of augmentation strategy is validated by ablation studies. This work offers a precise and effective solution for conservation and underwater robotics applications.

Figures

Figures reproduced from arXiv: 2506.23505 by the authors.

Figure 1
Figure 1. The overall architecture of the YOLOv12 model used in this study, illustrating the flow from the input image through the Backbone, Neck, and to the final detection output. Residual Efficient Layer Aggregation (R-ELAN) The R-ELAN block, illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Architectural diagram of the Residual Efficient Layer Aggregation (R-ELAN) block combining multi-path feature aggregation using A2 with residual connectivity to enhance gradient flow and feature robustness. Area Attention - A2 Underwater object detection faces persistent challenges with occlusion and turbidity that degrade conventional attention mechanisms. A2 addresses these limitations through a multi-head attenti… view at source ↗
Figure 3
Figure 3. Area Attention (A2) architecture showing the multi-head attention mechanism with spatial partitioning. The input feature map [B, C, H, W] is transformed into Key, Query, and Value representations with reduced spatial dimensions [B,(W ×H)/area, C] through area-based partitioning. The multi-head attention operates on these parti￾tioned features before reshaping back to original spatial dimensions [B, C, H, W] for outp… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visual comparison of physics-informed augmentation techniques applied to un￾derwater imagery: (a) Original benthic scene showing natural coloration and visibility; (b) HSV transformations simulating wavelength-dependent attenuation and turbidity effects; (c) Random era…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 31 canonical work pages

  1. [1]

    N. C. Thompson, K. Greenewald, K. Lee, and G. F. Manso,The computa- tional limits of deep learning, 2022. arXiv:2007.05558 [cs.LG]. [Online]. Available: https://arxiv.org/abs/2007.05558

    arXiv 2022

  2. [2]

    Self-attention and long-range relationship capture network for underwater object detection,

    Z. Gao, Y. Shi, and S. Li, “Self-attention and long-range relationship capture network for underwater object detection,”Journal of King Saud University-Computer and Information Sciences, vol. 36, p. 101971, 2024. 16 Tinh Nguyen

    work page 2024

  3. [3]

    An improved yolov5-based underwater object-detection framework,

    J. Zhang, J. Zhang, K. Zhou, Y. Zhang, H. Chen, and X. Yan, “An improved yolov5-based underwater object-detection framework,”Sensors, vol. 23, 2023,issn: 1424-8220. doi: 10.3390/s23073693. [Online]. Avail- able: https://www.mdpi.com/1424-8220/23/7/3693

    work page doi:10.3390/s23073693 2023

  4. [4]

    Two-stage underwater object detec- tion network using swin transformer,

    J. Liu, S. Liu, S. Xu, and C. Zhou, “Two-stage underwater object detec- tion network using swin transformer,”IEEE Access, vol. 10, pp. 117235– 117247, 2022. doi: 10.1109/ACCESS.2022.3219592

    arXiv 2022

  5. [5]

    Underwater object detection method based on improved faster rcnn,

    H. Wang and N. Xiao, “Underwater object detection method based on improved faster rcnn,” Applied Sciences, vol. 13, 2023, issn: 2076-3417. doi: 10 . 3390 / app13042746. [Online]. Available: https : / / www . mdpi . com/2076-3417/13/4/2746

    work page 2023

  6. [6]

    Yolo-dafs: A composite-enhanced un- derwater object detection algorithm,

    S. Luo, C. Dong, G. Dong, et al., “Yolo-dafs: A composite-enhanced un- derwater object detection algorithm,”Journal of Marine Science and En- gineering, vol. 13, 2025, issn: 2077-1312. doi: 10 . 3390 / jmse13050947. [Online]. Available:https://www.mdpi.com/2077-1312/13/5/947

    work page 2025

  7. [7]

    An improved yolov9s algorithm for underwater object detection,

    S. Zhou, L. Wang, Z. Chen, H. Zheng, Z. Lin, and L. He, “An improved yolov9s algorithm for underwater object detection,” Journal of Marine Science and Engineering, vol. 13, 2025, issn: 2077-1312. doi: 10.3390/ jmse13020230. [Online]. Available:https://www.mdpi.com/2077-1312/ 13/2/230

    work page 2025

  8. [8]

    Bi2f-yolo: A novel framework for underwater object detection based on yolov7,

    X. Liu, K. Zhao, C. Liu, and L. Chen, “Bi2f-yolo: A novel framework for underwater object detection based on yolov7,”Intelligent Marine Technol- ogy and Systems, vol. 3, Mar. 2025.doi: 10.1007/s44295-025-00060-9

    work page doi:10.1007/s44295-025-00060-9 2025

Show all 48 references
  1. [9]

    Yolov7-chs: An emerging model for underwater object detection,

    L. Zhao, Q. Yun, F. Yuan, X. Ren, J. Jin, and X. Zhu, “Yolov7-chs: An emerging model for underwater object detection,”Journal of Marine Science and Engineering, vol. 11, 2023, issn: 2077-1312. doi: 10.3390/ jmse11101949. [Online]. Available:https://www.mdpi.com/2077-1312/ 11/10/1949

    2023

  2. [10]

    Y. Tian, Q. Ye, and D. Doermann, Yolov12: Attention-centric real-time object detectors, 2025. arXiv: 2502 . 12524 [cs.CV]. [Online]. Available: https://arxiv.org/abs/2502.12524

    2025 arXiv

  3. [11]

    T. Dao, D. Y. Fu, S. Ermon, A. Rudra, and C. Ré,Flashattention: Fast and memory-efficient exact attention with io-awareness, 2022. arXiv: 2205 . 14135 [cs.LG] . [Online]. Available: https : / / arxiv . org / abs / 2205 . 14135

    2022

  4. [12]

    Huang, V

    J. Huang, V. Rathod, C. Sun,et al., Speed/accuracy trade-offs for modern convolutional object detectors, 2017. arXiv:1611.10012 [cs.CV]. [Online]. Available: https://arxiv.org/abs/1611.10012

    2017 arXiv

  5. [13]

    Girshick, J

    R. Girshick, J. Donahue, T. Darrell, and J. Malik, Rich feature hierar- chies for accurate object detection and semantic segmentation, 2014. arXiv: 1311.2524 [cs.CV]. [Online]. Available:https://arxiv.org/abs/1311. 2524

    2014 arXiv

  6. [14]

    Fast r-cnn,

    R. Girshick, “Fast r-cnn,” in2015 IEEE International Conference on Com- puter Vision (ICCV), 2015, pp. 1440–1448.doi: 10.1109/ICCV.2015.169. SignBart 17

    2015 doi

  7. [15]

    S. Ren, K. He, R. Girshick, and J. Sun,Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. arXiv: 1506.01497 [cs.CV]. [Online]. Available:https://arxiv.org/abs/1506.01497

    2016 arXiv

  8. [16]

    Redmon, S

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi,You only look once: Unified, real-time object detection, 2016. arXiv:1506.02640 [cs.CV]. [On- line]. Available:https://arxiv.org/abs/1506.02640

    2016 arXiv

  9. [17]

    Ssd: Single shot multibox detec- tor,

    W. Liu, D. Anguelov, D. Erhan,et al., “Ssd: Single shot multibox detec- tor,” inComputer Vision – ECCV 2016. Springer International Publishing, 2016, pp. 21–37,isbn: 9783319464480. doi: 10.1007/978-3-319-46448- 0 _ 2. [Online]. Available: http : / / dx . doi . org / 10 . 1007 ...

    2016 doi

  10. [18]

    Z. Li, L. Yang, and F. Zhou, Fssd: Feature fusion single shot multibox detector, 2024. arXiv: 1712.00960 [cs.CV] . [Online]. Available:https: //arxiv.org/abs/1712.00960

    2024 arXiv

  11. [19]

    Object detection system based on ssd algorithm,

    Q. Shuai and X. Wu, “Object detection system based on ssd algorithm,” in 2020 international conference on culture-oriented science & technology (ICCST), IEEE, 2020, pp. 141–144

    2020

  12. [20]

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár,Focal loss for dense object detection, 2018. arXiv: 1708 . 02002 [cs.CV]. [Online]. Available: https://arxiv.org/abs/1708.02002

    2018 arXiv

  13. [21]

    arXiv: 2005.12872 [cs.CV]

    N.Carion,F.Massa,G.Synnaeve,N.Usunier,A.Kirillov,andS.Zagoruyko, End-to-end object detection with transformers, 2020. arXiv: 2005.12872 [cs.CV]. [Online]. Available:https://arxiv.org/abs/2005.12872

    2020 arXiv

  14. [22]

    L. Chen, Y. Huang, J. Dong, et al., Underwater object detection in the era of artificial intelligence: Current, challenge, and future, 2024. arXiv: 2410 . 05577 [cs.CV]. [Online]. Available: https : / / arxiv . org / abs / 2410.05577

    2024 arXiv

  15. [23]

    Variational image dehazing with a novel underwater dark channel prior,

    Z. Jin, Y. Ma, L. Min, and M. Zheng, “Variational image dehazing with a novel underwater dark channel prior,” 2, vol. 19, 2025, pp. 334–354.doi: 10.3934/ipi.2024035. [Online]. Available:https://www.aimsciences. org/article/id/66cd5261e7a25d6c964e7386

    2025 doi

  16. [24]

    Underwater image enhancement of ROV usingmodifiedWaterNet,

    S. Wu, Y. Chen, and X. Yang, “Underwater image enhancement of ROV usingmodifiedWaterNet,” in Third International Symposium on Computer Applications and Information Systems (ISCAIS 2024), H. Wang and W. Li, Eds., ser. Society of Photo-Optical Instrumentation Engineers (SPIE) Co...

    2024

  17. [25]

    K. K. Babu, A. Tabassum, B. Navaneeth, T. Jahnavi, and Y. Akshaya, Underwater image enhancement using generative adversarial networks: A survey, 2025. arXiv: 2501.06273 [eess.IV] . [Online]. Available:https: //arxiv.org/abs/2501.06273

    2025 arXiv

  18. [26]

    An unsupervised underwater image en- hancement method based on generative adversarial networks with edge extraction,

    Y. Jia, Z. Wang, and L. Zhao, “An unsupervised underwater image en- hancement method based on generative adversarial networks with edge extraction,” Frontiers in Marine Science, vol. Volume 11 - 2024, 2024, issn: 2296-7745. doi: 10.3389/fmars.2024.1471014. [Online]. Available:...

    2024

  19. [27]

    New underwater image enhancement algorithm based on improved u-net,

    S. Zhu, Z. Geng, Y. Xie, et al., “New underwater image enhancement algorithm based on improved u-net,”Water, vol. 17, 2025,issn: 2073-4441. doi: 10.3390/w17060808. [Online]. Available:https://www.mdpi.com/ 2073-4441/17/6/808

    2025 doi

  20. [28]

    Yolov5-based enhanced underwater seaweed detection using open-source datasets,

    G. Prabhakar, N. V, S. S, and B. B, “Yolov5-based enhanced underwater seaweed detection using open-source datasets,”Recent Patents on Engi- neering,vol.19,May2024. doi: 10.2174/0118722121303581240527070227

    doi

  21. [29]

    Feb-yolov8: A multi-scale lightweight detec- tion model for underwater object detection,

    Y. Zhao, F. Sun, and X. Wu, “Feb-yolov8: A multi-scale lightweight detec- tion model for underwater object detection,”PLOS ONE, vol. 19, pp. 1– 19, Sep. 2024.doi: 10.1371/journal.pone.0311173. [Online]. Available: https://doi.org/10.1371/journal.pone.0311173

    2024 doi

  22. [30]

    Cstc-yolov8:Underwaterobject detection model based on improved yolov8 for side scan sonar images,

    Y.Luo,G.Huo,Z.Cheng,andW.Zhang,“Cstc-yolov8:Underwaterobject detection model based on improved yolov8 for side scan sonar images,” Journal of Electronic Imaging, vol. 34, p. 023019, 2025.doi: 10.1117/1. JEI.34.2.023019 . [Online]. Available:https://doi.org/10.1117/1. JEI.34.2.023019

    2025 doi

  23. [31]

    You only look once: Unified,real-timeobjectdetection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified,real-timeobjectdetection,” Jun.2016,pp.779–788. doi: 10.1109/ CVPR.2016.91

    2016

  24. [32]

    J. Dong, W. Wu, J. Cheng, and X. Tang,You sense only once beneath: Ultra-light real-time underwater object detection, 2025. arXiv:2504.15694 [cs.CV]. [Online]. Available:https://arxiv.org/abs/2504.15694

    2025

  25. [33]

    Refining features for underwater object detection at the frequency level,

    W. Wang, Z. Yu, and M. Huang, “Refining features for underwater object detection at the frequency level,”Frontiers in Marine Science, vol. Volume 12 - 2025, 2025, issn: 2296-7745. doi: 10 . 3389 / fmars . 2025 . 1544839. [Online]. Available:https://www.frontiersin.org/journals...

    2025

  26. [34]

    A new dataset, poisson gan and aquanet for underwater object grabbing,

    C. Liu, Z. Wang, S. Wang,et al., “A new dataset, poisson gan and aquanet for underwater object grabbing,”IEEE Transactions on Circuits and Sys- tems for Video Technology, vol. 32, pp. 2831–2844, 2021

    2021

  27. [35]

    Detectionofmarineanimalsinanewunderwaterdatasetwithvaryingvis- ibility,

    M. Pedersen, J. B. Haurum, R. Gade, T. B. Moeslund, and N. Madsen, “Detectionofmarineanimalsinanewunderwaterdatasetwithvaryingvis- ibility,” inThe IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR) Workshops, Jun. 2019

    2019

  28. [36]

    A dataset and benchmark of underwater object detection for robot picking,

    C. Liu, H. Li, S. Wang,et al., “A dataset and benchmark of underwater object detection for robot picking,” in2021 IEEE International Conference on Multimedia &; Expo Workshops (ICMEW), IEEE, Jul. 2021, pp. 1–6. doi: 10 . 1109 / icmew53276 . 2021 . 9455997. [Online]. Available:...

    2021

  29. [37]

    Scr-net: A novel lightweight aquatic biological detection network,

    T. Li, Y. Gang, S. Li, and Y. Shang, “Scr-net: A novel lightweight aquatic biological detection network,” PLOS ONE, vol. 20, pp. 1–25, Jun. 2025. doi: 10.1371/journal.pone.0324067. [Online]. Available:https://doi. org/10.1371/journal.pone.0324067. SignBart 19

    2025 doi

  30. [38]

    Lfn-yolo: Precision underwater small object detection via a lightweight reparameterized approach,

    M. Liu, Y. Wu, R. Li, and C. Lin, “Lfn-yolo: Precision underwater small object detection via a lightweight reparameterized approach,” Frontiers in Marine Science, vol. Volume 11 - 2024, 2025, issn: 2296-7745. doi: 10 . 3389 / fmars . 2024 . 1513740. [Online]. Available: https ...

    2024

  31. [39]

    Underwa- ter object classification and detection: First results and open challenges,

    A. Jesus, C. Zito, C. Tortorici, E. Roura, and G. De Masi, “Underwa- ter object classification and detection: First results and open challenges,” OCEANS 2022-chennai, pp. 1–6, 2022

    2022

  32. [40]

    Lightweight underwa- ter object detection based on yolo v4 and multi-scale attentional feature fusion,

    M. Zhang, S. Xu, W. Song, Q. He, and Q. Wei, “Lightweight underwa- ter object detection based on yolo v4 and multi-scale attentional feature fusion,” Remote Sensing, vol. 13, 2021, issn: 2072-4292. doi: 10. 3390 / rs13224706. [Online]. Available: https://www.mdpi.com/2072- 429...

    2021

  33. [41]

    Yolov8-mu: An improved yolov8 underwater detector based on a large kernel block and a multi-branch reparameterization module,

    X. Jiang, X. Zhuang, J. Chen, J. Zhang, and Y. Zhang, “Yolov8-mu: An improved yolov8 underwater detector based on a large kernel block and a multi-branch reparameterization module,”Sensors, vol. 24, 2024,issn: 1424-8220. doi: 10.3390/s24092905. [Online]. Available:https://www....

    2024 doi

  34. [42]

    C. Li, W. Liu, G. Gong, X. Ding, and X. Zhong,Su-yolo: Spiking neu- ral network for efficient underwater object detection, Mar. 2025.doi: 10. 48550/arXiv.2503.24389

    2025 doi arXiv

  35. [43]

    Jiang, X

    X. Jiang, X. Zhuang, J. Chen, and J. Zhang, Epbc-yolov8: An efficient and accurate improved yolov8 underwater detector based on an attention mechanism, 2025. arXiv:2502.05788 [cs.CV]. [Online]. Available:https: //arxiv.org/abs/2502.05788

    2025 arXiv

  36. [44]

    Vanilla-Yolo: a lightweight underwater object detector via reparameterization and multi-scale feature fusion,

    Y. Qi and J. Sun, “Vanilla-Yolo: a lightweight underwater object detector via reparameterization and multi-scale feature fusion,” in International Conference on Computer Application and Information Security (ICCAIS 2024), S. A. Safaa, P. H. Mohan, and B. Farid, Eds., Internati...

    2024 doi

  37. [45]

    Multi-scale feature enhancement method for underwater object detection,

    M. Li, W. Liu, C. Shao, B. Qin, A. Tian, and H. Yu, “Multi-scale feature enhancement method for underwater object detection,”Symmetry, vol. 17, 2025, issn: 2073-8994. doi: 10 . 3390 / sym17010063. [Online]. Available: https://www.mdpi.com/2073-8994/17/1/63

    2025

  38. [46]

    U-decn: End-to-end underwater object detec- tionconvnetwithimproveddenoisingtraining,

    Z. Liu, B. Wang, and Y. Li, “U-decn: End-to-end underwater object detec- tionconvnetwithimproveddenoisingtraining,” arXiv preprint arXiv:2408.05780, 2024

    2024

  39. [47]

    Mas-yolov11: An improved underwater object detection algorithm based on yolov11,

    Y. Luo, A. Wu, and Q. Fu, “Mas-yolov11: An improved underwater object detection algorithm based on yolov11,”Sensors, vol. 25, 2025,issn: 1424-

    2025

  40. [8220]

    [Online]

    doi: 10.3390/s25113433. [Online]. Available:https://www.mdpi. com/1424-8220/25/11/3433

    doi

Pith tools

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