Pith. sign in

REVIEW 3 major objections 4 minor 33 references

A Deep Learning Framework for Two-Dimensional, Multi-Frequency Propagation Factor Estimation

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read One network maps atmospheric refractivity to two-band radar propagation

desk verdict Useful single-frequency propagation-factor emulator, but the multi-frequency claim is ill-posed: the network gets no frequency input, so it cannot represent the M-to-F mapping for two bands. read the letter →

arxiv 2505.15802 v2 pith:RT2P3PY5 submitted 2025-05-21 cs.LG eess.SPphysics.ao-ph

Sarah E. Wessinger , Leslie N. Smith , Jacob Gull , Jonathan Gehman , Zachary Beever , Andrew J. Kammerer This is my paper · ORCID
classification cs.LGeess.SPphysics.ao-ph
keywords patternpropagationfactormodifiedrefractivitydeepneuralnetworksimage-to-imagetranslationmulti-frequencymarineatmosphericboundarylayerlossU-Net
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

RefractNet is an image-to-image neural network that takes a range-homogeneous vertical profile of modified refractivity $M(z)$ and produces the two-dimensional pattern propagation factor $F$ over range and altitude. The paper's central claim is that one such network can serve both 3 GHz (S-band) and 10 GHz (X-band) simultaneously, with test metrics close to networks trained for one frequency alone. This matters because standard parabolic-equation solvers are accurate but slow, so a fast trained surrogate could make propagation awareness practical in time-critical radar settings. The study also finds that training on $F$ and converting to decibel form $F_{\rm dB}$ afterward gives better structural fidelity than training on $F_{\rm dB}$ directly, and that higher-altitude domains yield better predictions because they resolve interference patterns fully.

What carries the argument

The central object is RefractNet, a UVCGANv2 image-to-image generator (a U-Net with a visual transformer at the bottleneck) that learns the mapping from an $M$-domain input image to an $F$-domain output image. The input is a modified refractivity profile interpolated to 256 altitude points and repeated across 256 range pixels, making a range-homogeneous image; the output is the corresponding 256 by 256 pattern propagation factor field. Supervised training with an equal combination of MSE and SSIM loss drives the mapping, with TEMPER supplying the ground-truth $F$ fields. The architecture comparison in the paper is what justifies choosing UVCGANv2 as RefractNet over simpler U-Nets, recursive networks, and transformer-based variants.

What would settle it

Run TEMPER for S-band at 3 GHz over a 0-30 m output domain using (a) the full 0-300 m modified refractivity profile and (b) the same profile truncated to 0-30 m; if the two $F$ fields differ materially in cases with elevated ducts or layers, the 30 m experiments do not test the claimed $M$-to-$F$ mapping. A simpler check would compare RefractNet's 30 m S-band predictions against TEMPER outputs for a synthetic profile with a strong elevated duct above 30 m.

Watch

Extended reading notes

Core claim

The paper establishes that deep neural networks can be trained to map modified refractivity to the pattern propagation factor over the entire range-altitude domain, and that a single network can account for two frequencies at once. Using TEMPER-generated ground truth from 70,322 training cases (half S-band, half X-band), RefractNet—a UVCGANv2-based image-to-image generator—was trained with a combined MSE and SSIM loss. Across the eight experiments, dual-frequency test metrics were statistically comparable to single-frequency training: X-band-only experiments performed best overall, while dual-frequency predictions for each individual frequency landed close to the single-frequency baselines. The paper therefore claims a practical alternative to repeated parabolic equation simulation, with the caveat that results are sensitive to training choices such as target variable, altitude range, and frequency mix.

Load-bearing premise

The 30 m dual-frequency experiments assume that S-band propagation factor below 30 m is fully determined by the modified refractivity profile below 30 m, so truncating the input to 30 m does not discard information that TEMPER's output depends on.

Editorial extensions

If this is right

  • If the claim holds, a single trained RefractNet can replace TEMPER for both S- and X-band propagation factor estimation on the tested scenarios, cutting per-case compute time after training.
  • Training on $F$ rather than $F_{\rm dB}$ is the recommended route; converting predicted $F$ to $F_{\rm dB}$ and propagation loss preserves more of the interference structure.
  • Dual-frequency training is almost as accurate as single-frequency training, so a multi-frequency emulator does not force a large accuracy penalty.
  • Higher-altitude evaluation domains (0-300 m) are preferable to 0-30 m windows because they fully resolve constructive interference patterns; future work should choose evaluation domains with this in mind.
  • Results are sensitive to how training is set up—target variable, altitude, single versus multiple frequencies—so an emulator must be designed around its intended operational domain.

Reading between the lines

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

  • A natural extension the paper does not test is conditioning the network on frequency as an additional input channel, which could let one trained model serve arbitrary frequencies rather than a fixed pair.
  • Because the input profile is range-homogeneous, the emulator is only validated for horizontally uniform environments; feeding the network a two-dimensional $M$ field with range-varying structure would test whether the mapping generalizes to non-homogeneous ducts.
  • The 30 m S-band experiments implicitly assume that propagation in the lowest 30 m is determined by refractivity in those same 30 m; comparing TEMPER runs with truncated versus full 0-300 m profiles would show whether elevated structures above 30 m matter.
  • The network inherits any bias in TEMPER's physics, such as the smooth-sea assumption, so its predictions are only as good as the simulator it learned from.
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

3 major / 4 minor

Summary. This paper presents RefractNet, a UVCGANv2-based image-to-image generator that takes a range-homogeneous modified refractivity (M) image and predicts a two-dimensional pattern propagation factor (F) image over range and altitude. The authors train and test on TEMPER simulations at 3 and 10 GHz, compare single-frequency and dual-frequency training for F and FdB outputs at 0-30 m and 0-300 m altitudes, and report MSE, SSIM, and FID metrics. The central claim is that a single network can account for multiple frequencies while achieving held-out performance comparable to single-frequency networks.

Significance. If the multi-frequency claim were supported, RefractNet would be a fast surrogate for TEMPER on range-homogeneous profiles, with practical value for radar propagation assessment. The paper's strengths are its well-structured experimental matrix (Table II), transparent hyperparameters (Table I), held-out test sets, and comparison across several generator architectures. The main limitation is that the described architecture has no frequency conditioning, so the multi-frequency claim is not currently supported. The paper also provides no code or data release, which limits reproducibility.

major comments (3)
  1. [Section II.A and Section III.A] The dual-frequency Experiments 1 and 4 are described as training a single network on both X-band and S-band cases, but the network input is only a 256x256 image of range-homogeneous M(z) with no frequency channel, embedding, or other conditioning mechanism. TEMPER's F is frequency-dependent: the same M profile at 3 and 10 GHz produces different F images. A deterministic feed-forward network with only M as input cannot represent this one-to-many mapping. If the same M appears with both frequency labels in training, the loss is minimized by an average prediction that cannot match either target; if each M is paired with only one frequency, the network cannot transfer to the other frequency for a new M. The per-frequency metrics in Table V therefore do not demonstrate multi-frequency capability; they only show performance on two subsets of inputs that happen to come from different frequencies. The authors should either add explicit frequency conditioning or revise the central claim to single-frequency emulation.
  2. [Section III.B and Table III] All 30 m experiments (Experiments 1-6) use M(z) only over 0-30 m, repeated across range, while the S-band TEMPER runs described in Section II.A natively cover 0-300 m. The paper does not justify that S-band F in 0-30 m is insensitive to M above 30 m. Elevated ducts or other layers above 30 m can influence low-altitude propagation at 3 GHz, so the 30 m S-band results may reflect an artificially limited input domain. This concern affects Experiments 1, 3, 4, and 6, and should be addressed by a sensitivity analysis or by using the full M profile as input.
  3. [Section III.B] All quantitative claims are based on metrics computed on min-max normalized images, as defined in Table III and Equation 5. The reported MSE values are therefore not directly interpretable in physical units of F or FdB. For example, a normalized MSE of 0.001 at 30 m F corresponds to a different physical error than the same value at 300 m F. The authors should report denormalized errors (e.g., in dB or linear F) to make the engineering accuracy concrete.
minor comments (4)
  1. [Section III.B] FID is a distributional metric defined over a set of real and generated samples; computing FID separately for each test image and then averaging is not a valid use of the metric. The authors should either compute FID over batches of test cases or rely on SSIM and MSE for per-sample evaluation.
  2. [Section III.B] The t-tests compare populations of about 15,000 samples, so p-values below 0.05 can reflect negligible effect sizes. Reporting effect sizes or confidence intervals would make the statistical comparisons more informative.
  3. [Section II.A] The comparison of performance between F and FdB is confounded by the different normalization schemes in Table III; normalized MSE values for the two variables are not directly comparable.
  4. [Section I.A] The phrase 'through For PL' in the introduction appears to be a typographical error, and reference [2] is missing from the reference list.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: RefractNet is a supervised surrogate validated on held-out TEMPER outputs, which is standard emulation practice, not an input-derived prediction.

full rationale

RefractNet is trained on (M, F) pairs generated by the TEMPER propagation solver and evaluated on held-out test cases from the same simulator. This is a conventional emulator protocol: the test cases are disjoint from the training cases, so the reported MSE, SSIM, and FID values measure generalization to unseen M profiles rather than reconstruction of the training labels. The paper does not fit a parameter to the test outputs and then report those same outputs as predictions. The dual-frequency experiments train one network on a mixed S-band/X-band dataset and compare per-frequency metrics to single-frequency models; the lack of an explicit frequency conditioning channel is a legitimate experimental-design and ill-posedness concern about whether the network can represent a one-to-many M-to-F mapping, but that concern does not make the claim circular. The conclusion that the network can 'reasonably predict' F does not reduce by construction to the training inputs, nor does it depend on a load-bearing self-citation: UVCGANv2 is an externally cited architecture, and renaming it 'RefractNet' is a terminological choice, not an ansatz imported from the authors' prior work. No uniqueness theorem or circular definition is invoked. Therefore, under the quoted-evidence standard, no circular step is present.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper contributes no new physical entities. Its central claim rests on restrictive domain assumptions: range-homogeneous refractivity, smooth sea surface, fixed antenna height, TEMPER as truth, and truncated S-band inputs in the 30 m experiments. The fitted quantities are network weights, architecture hyperparameters, and normalization constants, not physical constants.

free parameters (5)
  • Learned generator weights = Learned during training on TEMPER targets
    The M-to-F mapping is entirely empirical; no physics is imposed, so the central result depends on these fitted weights.
  • UVCGANv2 block count = 12
    Architecture depth chosen by Optuna Bayesian optimization (Table I); affects model capacity and results.
  • UVCGANv2 feature count = 384
    Feature width chosen by Optuna; affects model capacity.
  • MSE and SSIM loss weights = 0.5 and 0.5
    The loss is an equal combination of MSE and SSIM; the authors tested variations and settled on this but report no alternatives.
  • Min-max normalization extrema Vmin and Vmax = Varies by altitude and variable (Table III)
    Data-dependent normalization constants set the scale of all reported MSE values, for example F is divided by 16.45.
assumptions (5)
  • domain assumption M is horizontally homogeneous over range and time
    The 2D input image is a single vertical M profile repeated across 256 range bins (Section I.A and Figure 1), so all results are limited to range-homogeneous environments.
  • domain assumption TEMPER parabolic equation output is the reference definition of F
    All training labels and all evaluation metrics use TEMPER version 3.2.3 (Section II.A); no measured propagation factor validates the surrogate.
  • domain assumption Smooth sea surface is sufficient
    TEMPER runs assume a smooth sea surface, so rough-ocean effects on F are outside the scope (Section II.A).
  • domain assumption Fixed 20 m antenna height
    Every simulation uses a 20 m antenna height (Section II.A), so the network is trained only for this geometry.
  • ad hoc to paper The 0-30 m M profile is sufficient to predict S-band F below 30 m in the 30 m experiments
    S-band TEMPER natively covers 0-300 m but the 30 m experiments truncate input and output to 0-30 m (Sections II.A and III.A). If elevated layers matter, this premise fails.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Deep Learning Framework for Two-Dimensional, Multi-Frequency Propagation Factor Estimation." pith.science (2026). https://pith.science/paper/RT2P3PY5

@misc{pith2026250515802,
  author       = {Pith},
  title        = {Pith review of: A Deep Learning Framework for Two-Dimensional, Multi-Frequency Propagation Factor Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RT2P3PY5}},
  note         = {Machine review of arXiv:2505.15802}
}
read the original abstract

Accurately estimating the refractive environment over multiple frequencies within the marine atmospheric boundary layer is crucial for the effective deployment of radar technologies. Traditional parabolic equation simulations, while effective, can be computationally expensive and time-intensive, limiting their practical application. This communication explores a novel approach using deep neural networks to estimate the pattern propagation factor, a critical parameter for characterizing environmental impacts on signal propagation. Image-to-image translation generators designed to ingest modified refractivity data and generate predictions of pattern propagation factors over the same domain were developed. Findings demonstrate that deep neural networks can be trained to analyze multiple frequencies and reasonably predict the pattern propagation factor, offering an alternative to traditional methods.

Figures

Figures reproduced from arXiv: 2505.15802 by the authors.

Figure 1
Figure 1. (a) M(z) used to create the (b) range-homogeneous M domain input for RefractNet. contain on the order of 50,000 training samples (i.e., MNIST, Cifar￾10/100). Note, M is assumed homogeneous over range and time for this study ( [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. F domain calculated by (a) TEMPER, (b) RefractNet, and (c) the difference between these domains, TEMPER-RefractNet, from a case in Experiment 1 where the evaluation metrics are: MSE = 1.086x10-5, FID = 0.0016, and SSIM = 0.930 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Distribution of evaluation metrics for SSIM (a, d, g, j), MSE (b, e, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: FdB domain calculated by (a) TEMPER, (b) RefractNet, and (c) the difference, TEMPER-RefractNet, between these domains for a case from Experiment 6 where the evaluation metrics are: MSE = 0.0020, FID = 0.515, and SSIM = 0.758 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: F domain for the same case in both (a) Experiment 1 with metrics of MSE = 1.086x10-5, FID = 0.0016, and SSIM = 0.930, and (b) Experiment 2 of MSE = 8.035x10-6, FID = 0.0012, and SSIM = 0.843 and (c) the difference between these domains [PITH_FULL_IMAGE:figures/full_fi…
Figure 6
Figure 6. Figure 6: FdB domain (a,d) from TEMPER, (b) calculated from Experiment 1 F (c) the difference between a and b, (e) from Experiment 4, and (f) the difference between d and e [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Distributions of (a) SSIM, (b) MSE, and (c) FID for Experiment 4 compared to the calculated [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: F domain calculated by (a) TEMPER, (b) RefractNet, and (c) the difference between these domains for a case from Experiment 7 where the evaluation metrics are: MSE = 0.0021, FID = 0.505, and SSIM = 0.749 [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 27 canonical work pages

  1. [1]

    A note on a simple transmission formula,

    H. T. Friis, “A note on a simple transmission formula,” Proceedings of the IRE, vol. 34, no. 5, pp. 254–256, 1946. [2]

    work page 1946

  2. [3]

    A terrain parabolic equation model for propagation in the troposphere,

    A. E. Barrios, “A terrain parabolic equation model for propagation in the troposphere,” IEEE Transactions on Antennas and Propagation, vol. 42, no. 1, pp. 90–98, 1994

    work page 1994

  3. [4]

    User’s guide for the vtrpe (variable terrain radio parabolic equation) computer model,

    F. J. Ryan, “User’s guide for the vtrpe (variable terrain radio parabolic equation) computer model,” Final Report, 1991

    work page 1991

  4. [5]

    Simulated microwave propagation through tropospheric turbulence,

    D. Rouseff, “Simulated microwave propagation through tropospheric turbulence,” IEEE transactions on antennas and propagation , vol. 40, no. 9, pp. 1076–1083, 1992

    work page 1992

  5. [6]

    Radio meteorology. national bureau of standards monogr,

    B. Bean and E. Dutton, “Radio meteorology. national bureau of standards monogr,” 1968

    work page 1968

  6. [7]

    Path loss prediction based on machine learning: Principle, method, and data expansion,

    Y . Zhang, J. Wen, G. Yang, Z. He, and J. Wang, “Path loss prediction based on machine learning: Principle, method, and data expansion,” Applied Sciences, vol. 9, no. 9, p. 1908, 2019

    work page 1908

  7. [8]

    Machine learning in the air,

    D. G ¨und¨uz, P. De Kerret, N. D. Sidiropoulos, D. Gesbert, C. R. Murthy, and M. Van der Schaar, “Machine learning in the air,” IEEE Journal on Selected Areas in Communications , vol. 37, no. 10, pp. 2184–2199, 2019

    work page 2019

  8. [9]

    Deep learning-based channel estimation for doubly selective fading channels,

    Y . Yang, F. Gao, X. Ma, and S. Zhang, “Deep learning-based channel estimation for doubly selective fading channels,” IEEE Access , vol. 7, pp. 36 579–36 589, 2019

    work page 2019

Show all 33 references
  1. [10]

    Regression of large-scale path loss parameters using deep neural networks,

    M. Bal, A. Marey, H. F. Ates, T. Baykas, and B. K. Gunturk, “Regression of large-scale path loss parameters using deep neural networks,” IEEE Antennas and Wireless Propagation Letters , vol. 21, no. 8, pp. 1562– 1566, 2022

    2022

  2. [11]

    Deepray: Deep learning meets ray-tracing,

    S. Bakirtzis, K. Qiu, J. Zhang, and I. Wassell, “Deepray: Deep learning meets ray-tracing,” in 2022 16th European Conference on Antennas and Propagation (EuCAP). IEEE, 2022, pp. 1–5

    2022

  3. [12]

    Generalizable physics-guided convolutional neural network for irregular terrain prop- agation,

    S. Huang, H. Qin, W. Hou, X. Zhang, and X. Zhang, “Generalizable physics-guided convolutional neural network for irregular terrain prop- agation,” IEEE Transactions on Antennas and Propagation , 2025

    2025

  4. [13]

    A study on radar target detection based on deep neural networks,

    L. Wang, J. Tang, and Q. Liao, “A study on radar target detection based on deep neural networks,” IEEE Sensors Letters , vol. 3, no. 3, pp. 1–4, 2019

    2019

  5. [14]

    An evaporation duct height prediction method based on deep learning,

    X. Zhu, J. Li, M. Zhu, Z. Jiang, and Y . Li, “An evaporation duct height prediction method based on deep learning,” IEEE Geoscience and Remote Sensing Letters , vol. 15, no. 9, pp. 1307–1311, 2018

    2018

  6. [15]

    Refractivity inversions from point-to-point x-band radar propagation measurements,

    D. M. Pastore, S. E. Wessinger, D. P. Greenway, M. J. Stanek, R. J. Burkholder, T. Haack, Q. Wang, and E. E. Hackett, “Refractivity inversions from point-to-point x-band radar propagation measurements,” Radio Science, vol. 57, no. 2, pp. 1–16, 2022

    2022

  7. [16]

    The com- parison of long short-term memory neural network and deep forest for the evaporation duct height prediction,

    Q. Liao, Y . Mai, Z. Sheng, Y . Wang, Q. Ni, and S. Zhou, “The com- parison of long short-term memory neural network and deep forest for the evaporation duct height prediction,” IEEE Transactions on Antennas and Propagation, vol. 71, no. 5, pp. 4444–4450, 2023

    2023

  8. [17]

    Path loss prediction in evaporation ducts based on deep neural network,

    B. Shu, W. Zhang, Y . Chen, J. Sun, and C.-X. Wang, “Path loss prediction in evaporation ducts based on deep neural network,” IEEE Antennas and Wireless Propagation Letters , 2023

    2023

  9. [18]

    Determination of neural network parameters for path loss prediction in very high frequency wireless channel,

    S. I. Popoola, A. Jefia, A. A. Atayero, O. Kingsley, N. Faruk, O. F. Oseni, and R. O. Abolade, “Determination of neural network parameters for path loss prediction in very high frequency wireless channel,” IEEE access, vol. 7, pp. 150 462–150 483, 2019

    2019

  10. [19]

    A deep neural network-based multi- frequency path loss prediction model from 0.8 ghz to 70 ghz,

    C. Nguyen and A. A. Cheema, “A deep neural network-based multi- frequency path loss prediction model from 0.8 ghz to 70 ghz,” Sensors, vol. 21, no. 15, p. 5100, 2021

    2021

  11. [20]

    Performance evaluation of machine learning methods for path loss prediction in rural environment at 3.7 ghz,

    N. Moraitis, L. Tsipi, D. V ouyioukas, A. Gkioni, and S. Louvros, “Performance evaluation of machine learning methods for path loss prediction in rural environment at 3.7 ghz,” Wireless Networks, vol. 27, no. 6, pp. 4169–4188, 2021

    2021

  12. [21]

    Deep learning method for path loss prediction in mobile communication sys- tems,

    Z. Xu, H. Cao, Y . Yin, X. Zhang, L. Wu, D. He, and Y . Wang, “Deep learning method for path loss prediction in mobile communication sys- tems,” in 2021 13th International Symposium on Antennas, Propagation and EM Theory (ISAPE) . IEEE, 2021, pp. 01–03

    2021

  13. [22]

    Image-to-image translation with conditional adversarial networks,

    P. Isola, J.-Y . Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1125– 1134

    2017

  14. [23]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2223–2232

    2017

  15. [24]

    Image-to-image translation: Methods and applications,

    Y . Pang, J. Lin, T. Qin, and Z. Chen, “Image-to-image translation: Methods and applications,” IEEE Transactions on Multimedia , vol. 24, pp. 3859–3881, 2021

    2021

  16. [25]

    Deep generative adversarial networks for image-to-image translation: A review,

    A. Alotaibi, “Deep generative adversarial networks for image-to-image translation: A review,” Symmetry, vol. 12, no. 10, p. 1705, 2020

    2020

  17. [26]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 ...

    2015

  18. [27]

    Efficientnet: Rethinking model scaling for con- volutional neural networks,

    M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for con- volutional neural networks,” in International conference on machine learning. PMLR, 2019, pp. 6105–6114

    2019

  19. [28]

    U-net using stacked di- lated convolutions for medical image segmentation,

    S. Wang, S.-Y . Hu, E. Cheah, X. Wang, J. Wang, L. Chen, M. Baikpour, A. Ozturk, Q. Li, S.-H. Chou et al. , “U-net using stacked di- lated convolutions for medical image segmentation,” arXiv preprint arXiv:2004.03466, 2020

    2004 arXiv

  20. [29]

    Rethinking cyclegan: Improving quality of gans for unpaired image-to-image translation,

    D. Torbunov, Y . Huang, H.-H. Tseng, H. Yu, J. Huang, S. Yoo, M. Lin, B. Viren, and Y . Ren, “Rethinking cyclegan: Improving quality of gans for unpaired image-to-image translation,” 2023

    2023

  21. [30]

    Uvcgan v2: An improved cycle-consistent gan for unpaired image-to-image translation,

    ——, “Uvcgan v2: An improved cycle-consistent gan for unpaired image-to-image translation,” arXiv preprint arXiv:2303.16280 , 2023

    2023 arXiv

  22. [31]

    Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with trans- former,

    H. Wang, P. Cao, J. Wang, and O. R. Zaiane, “Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with trans- former,” in Proceedings of the AAAI conference on artificial intelligence, vol. 36, no. 3, 2022, pp. 2441–2449

    2022

  23. [32]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE transactions on image processing , vol. 13, no. 4, pp. 600–612, 2004

    2004

  24. [33]

    Frechet inception distance (fid) for evaluating gans,

    Y . Yu, W. Zhang, and Y . Deng, “Frechet inception distance (fid) for evaluating gans,” China University of Mining Technology Beijing Graduate School, vol. 3, no. 11, 2021

    2021

  25. [34]

    Global sensitivity of parabolic equation radar wave propagation simulation to sea state and atmospheric refractivity structure,

    N. Lentini and E. Hackett, “Global sensitivity of parabolic equation radar wave propagation simulation to sea state and atmospheric refractivity structure,” Radio Science, vol. 50, no. 10, pp. 1027–1049, 2015

    2015

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.