Pith. sign in

REVIEW 5 major objections 6 minor 29 references

A parallel auxiliary encoder that feeds input-derived features into every ResNet stage lifts license plate denoising and deblurring, reaching 31.052 dB PSNR with just ~1% more parameters.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-04 20:40 UTC pith:7FXYNBXE

load-bearing objection A plausible but sloppily presented architecture idea whose central comparison is not reproducible from the text—worth a round of revision, not acceptance as-is. the 5 major comments →

arxiv 2509.08392 v2 pith:7FXYNBXE submitted 2025-09-10 cs.CV

VRAE: Vertical Residual Autoencoder for License Plate Denoising and Deblurring

classification cs.CV
keywords license plate denoisinglicense plate deblurringvertical residual autoencoderauxiliary feature injectionimage restorationentropy preservationreal-time enhancementtraffic surveillance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper is trying to establish that a license-plate image enhancer can get autoencoder-level speed and parameter efficiency without sacrificing detail by adding a parallel auxiliary encoder that injects input-derived features into every stage of a ResNet backbone. If true, it would give real-time traffic surveillance systems a cheap pre-processing step that restores plates before recognition, where GANs are too slow or hallucinate digits and flow models are too heavy. The authors report that the best variant, VRAE3, reaches 31.052 dB PSNR and 0.898 SSIM, beating same-depth autoencoders, GANs, and flow-based models; they also show VRAE models sit on the Pareto front of quality versus speed. The key claim is that preserving entropy in early encoder layers prevents premature information collapse and leads to smoother, more stable compression.

Core claim

VRAE claims that a simple architectural addition—a parallel auxiliary encoder that injects features extracted directly from the degraded input into each stage of a ResNet-50 main encoder via element-wise addition—consistently improves reconstruction quality over plain autoencoders, GANs, and flow-based models on synthetically degraded vehicle images. In the paper's experiments, VRAE3 reaches 31.052 dB PSNR and 0.898 SSIM, the best in its comparison, while VRAE2 reaches 399 FPS; at equal stack depth, VRAE improves PSNR by about 20%, cuts NMSE roughly in half, and raises SSIM about 1% with only ~1% more parameters. The paper attributes the gain to better preservation of information in early en

What carries the argument

The vertical residual auxiliary block: a parallel stack of shallow Conv–Norm–Activation encoders that take the raw degraded image x and produce feature maps x'_i, added element-wise to the main encoder stage output before the next ResNet stage. This carries the argument because it is the only difference between VRAE and the AE baselines; it injects input-aware features at every stage without increasing feature dimensionality, so parameter overhead stays near 1% while gradient flow and early-layer information retention improve.

Load-bearing premise

The synthetic degradation model (uniform noise scaled by 0.1 plus ten passes of 3x3 average pooling) is assumed to stand in for real surveillance noise and motion blur; the paper provides no real-world validation, so the reported gains may not carry over to actual plate images.

What would settle it

Train VRAE3 and its strongest same-depth autoencoder on real low-light, motion-blurred traffic frames with known plate labels, then compare end-to-end license plate recognition accuracy; if the PSNR/SSIM advantage does not translate into recognition gains, or if the advantage disappears when using Gaussian blur or sensor noise rather than the paper's pooling-plus-uniform-noise degradation, the central claim is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Same-depth comparison: VRAE raises PSNR by about 20%, cuts NMSE by around 50%, and improves SSIM by about 1% for roughly 1% parameter overhead.
  • VRAE2 runs at 399 FPS, nearly matching AE2's 411 FPS, so the quality gain does not cost real-time operation.
  • VRAE3, the best configuration, reaches 31.052 dB PSNR and 0.898 SSIM while beating GAN- and flow-based baselines that have far more parameters.
  • The entropy analysis predicts that early-layer information retention is the mechanism, so the architecture is expected to stabilize quality across depths rather than only at one depth.
  • Deeper VRAE variants trade off parameter count and speed for quality, so the practical choice is VRAE2 or VRAE3.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A decisive next test is end-to-end license plate recognition on real surveillance footage; the paper's pixel metrics may not capture whether restoration actually helps a recognition model.
  • Because the auxiliary fusion is a generic additive feature-injection mechanism, it could transfer to other small-region restoration tasks such as face or QR-code enhancement, but this is an extension the paper does not test.
  • The entropy-preservation story could be tested directly by ablating the auxiliary block's injection at the first encoder stage only; the paper's own numbers suggest that is where the largest entropy difference appears.
  • If real-world degradations differ from the pooling-plus-noise model, the reported gains may shrink; training on a wider degradation distribution would be the natural robustness check.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The manuscript proposes a Vertical Residual Autoencoder (VRAE) for license plate image denoising and deblurring. The architecture is a ResNet-50-based autoencoder augmented with a parallel auxiliary branch that extracts features directly from the degraded input and fuses them by element-wise addition into the main encoder at each stage. The authors evaluate a family of VRAE models of depths 2–5 against same-depth autoencoders, a GAN, and a flow-based model on synthetically degraded images from the CCPD dataset, reporting PSNR, SSIM, NMSE, parameter counts, and FPS. They argue that the auxiliary input-feature injection stabilizes performance across depths and preserves information better than plain autoencoders, and they support this with an entropy-change analysis. The paper also includes a Pareto-front trade-off analysis and a limitations section.

Significance. If the central result is reproducible, the contribution is of moderate practical value: a lightweight architectural modification that improves restoration quality over same-depth autoencoders at less than 1% parameter overhead and with fast inference would be relevant for resource-constrained traffic surveillance. The paper supplies quantitative comparisons, parameter counts, and inference speed, which is a strength. However, the evaluation is entirely synthetic and the architecture is not described precisely enough to reproduce the reported numbers. The headline percentages in the abstract do not match Table 1 for most depths, and the degradation model contradicts the problem statement. These issues currently prevent verification of the paper's main claims.

major comments (5)
  1. [Secs. 2.1–2.3, Eqs. (1)–(2)] The notation for the auxiliary and main encoders is internally inconsistent. Sec. 2.1 defines the auxiliary encoders as {E_i} and the main encoders as {E′_i}, but Eq. (1) defines the auxiliary outputs as x′_i = E′_i(x), and Eq. (2) uses E_i as the main encoder. This makes the core architecture ambiguous. In addition, the fusion rule x_i = E_i(x_{i-1} + x′_{i-1}) requires x_{i-1} and x′_{i-1} to have the same spatial size and number of channels, yet the paper gives no downsampling/upsampling rule for the auxiliary features; Sec. 2.2 only says each block 'optionally includes pooling.' ResNet-50 stages change spatial resolution at specific points, so without an explicit matching mechanism the element-wise addition is undefined at typical stage boundaries. Please clarify the notation and specify the exact resolution/channel alignment for every fusion point.
  2. [Secs. 2 and 3.1] The degradation model is self-contradictory. The problem statement says 'Assume that all distortion is created by ϵ∼N(0,1)' and defines x = x̂ + ϵ. But Sec. 3.1 generates low-resolution inputs by 'adding discrete noise (values in [0,9] scaled by 0.1) and applying a 3×3 average pooling filter iteratively 10 times.' These are different corruption processes, and the blur component is absent from the problem statement. Also, '3×3 average pooling iteratively 10 times' is ambiguous: standard pooling with stride 2 or 3 would reduce the 256×256 image to a tiny or nonexistent spatial size, whereas stride 1 with padding is effectively a blur. Please state the exact noise distribution and the exact pooling strides/padding, and align the mathematical model with the implemented degradation.
  3. [Abstract and Sec. 4.1, Table 1] The headline improvement percentages are not supported by Table 1 for most depths. Claimed 'PSNR by about 20%': relative improvement from AE3 to VRAE3 is 31.052/26.159 ≈ 1.187 (18.7%), while depth 2 gives ≈ 9.1%, depth 4 ≈ 2.9%, and depth 5 ≈ 10.2%. Claimed 'SSIM by about 1%': VRAE3 vs AE3 is +0.096 absolute and +12% relative; the smallest relative increase is VRAE4 vs AE4 at +1.9%. Claimed 'NMSE by around 50%': the reductions are roughly 43% (depth 2), 73% (depth 3), 20% (depth 4), and 63% (depth 5). Also, expressing PSNR changes as percentages without a stated convention is nonstandard. The underlying claim that VRAE improves over AE at every depth is supported, but the specific quantitative summary needs correction and clarification.
  4. [Sec. 2.2 and Table 1] The reported parameter counts do not follow from the described architecture. Table 1 shows VRAE2 adds 1K parameters over AE2, VRAE3 adds 10K, VRAE4 adds 20K, and VRAE5 adds 50K. If each auxiliary block is a 'shallow Conv–Norm–Activation' block, the parameter increments should scale with the number of channels at each ResNet-50 stage; the nonlinear increments listed are not derivable from the text. Without a layer-by-layer description of each auxiliary block or access to code, the claimed 'marginal increase of roughly 1% in parameters' cannot be verified. Please provide the exact configuration of each auxiliary block or release the implementation.
  5. [Sec. 3.3 and Sec. 4.1, Table 1] No error bars or statistical tests are reported for the quantitative comparisons. The test set contains about 455 images (15% of 3,036), and several differences, such as AE4 PSNR 29.404 vs VRAE4 PSNR 30.246, or AE2 vs VRAE2, could plausibly fall within run-to-run variance. More importantly, all validation is on synthetically degraded images; no real traffic-surveillance images are used, so the paper's central practical claim that VRAE improves real license plate restoration is not demonstrated. Please report multiple runs or confidence intervals, and either add a real-world evaluation or substantially moderate the external-validity claims.
minor comments (6)
  1. [Title/Conclusion] The title and abstract call the method 'Vertical Residual Autoencoder,' but the conclusion calls it 'Variational Residual Autoencoder.' Please make the terminology consistent.
  2. [Sec. 1] There are typos and grammatical issues, e.g., 'Introdution & Related Work,' 'a fast realtime manner,' and several awkward phrasings. A light copyedit is recommended.
  3. [Sec. 4.2, Eq. (8)] The entropy-change proxy in Eq. (8) is imported from reference [29] but is not derived or adequately defined. The notation c_11 and the logarithm are unclear for zero or negative filter coefficients, and the extension from convolutional layers to entire encoder blocks is not justified. This analysis is secondary, but it should be made self-contained if it is retained.
  4. [Sec. 3.2] The GAN and FB baselines are described only as being 'designed with main encoder and decoder blocks consistent with our proposed method.' Please specify the generator/discriminator architecture, the flow architecture, and any loss weights other than α, so the comparison is reproducible.
  5. [Sec. 4.3, Fig. 3] The Pareto-front figure is described qualitatively. It would help to state which models are on the Pareto front in each of the three plots and to explain how the front is computed.
  6. [References] Reference [16] duplicates the content of [11] (same paper with slightly different page ranges), and reference [7] is cited for a convolutional encoder-decoder although it is a residual CNN denoiser. Please check the reference list for accuracy.

Circularity Check

0 steps flagged

No circular derivation; empirical gains are measured against baselines, and the entropy analysis is post-hoc interpretation using an external proxy.

full rationale

I walked the paper's derivation chain. The central claim—that VRAE outperforms AE/GAN/FB—is supported by Table 1, which reports measured PSNR/NMSE/SSIM/FPS values. No parameter was fitted to the test metrics and then renamed as a prediction. The architecture equations (1)-(3) define an encoder–auxiliary–decoder composition, Eq. (4) is a standard MSE loss, and nothing in Eqs. (1)-(9) makes the output equal to an input by construction. The entropy-preservation discussion (Sec. 4.2) is explicitly post-hoc ('consistent with our hypothesis'), and its proxy Eq. (8) is imported from external reference [29], not from the authors' own equations; the authors do not cite themselves anywhere. The inconsistencies flagged in the skeptic headline—symbol mismatch between E_i and E'_i, missing resolution-matching rules for element-wise addition, nonlinear parameter-count scaling, and abstract-vs-table numeric discrepancies—are reproducibility and correctness concerns, not circularity: they do not show that the reported improvements reduce to the inputs of the method. Under the hard rule that circularity must be exhibited by quoting a specific reduction, no circular step can be identified. Score 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

Everything the numerical claim rests on beyond the raw dataset: a hand-chosen synthetic degradation schedule, an imported entropy proxy from [29], an unverified initialization assumption, and an assumed equivalence between MSE minimization and recognizability.

free parameters (3)
  • noise amplitude range = 0 to 0.9 after scaling
    Hand-chosen in Sec 3.1 to simulate low-quality footage; affects all reported metrics.
  • blur iterations = 10
    Hand-chosen 3x3 average pooling repetitions in Sec 3.1; defines task difficulty.
  • GAN adversarial weight alpha = 1e-3
    Adopted from [21] in Sec 3.3; affects the GAN baseline only, not the VRAE claim.
axioms (4)
  • ad hoc to paper All distortion is produced by independent noise epsilon ~ N(0,1) added to clean image (Sec 2 Problem statement).
    Stated as the problem model, but the actual degradation in Sec 3.1 is uniform noise plus average pooling, so the assumption contradicts the experiment.
  • domain assumption Entropy-change proxy in Eq (8) from [29] holds for encoder blocks in this architecture.
    Sec 4.2 imports a proof from [29] without re-deriving it for block-level feature maps.
  • domain assumption ResNet-50 stage blocks are an appropriate main encoder for restoration; pretraining or initialization is unspecified.
    Sec 2.3; no initialization or pretraining details are given.
  • domain assumption MSE loss alone is sufficient for license plate denoising and deblurring.
    Sec 2.4 defines only MSE; no perceptual or recognition loss is used, yet the claim is about enhancing recognition.

pith-pipeline@v1.3.0-alltime-deepseek · 7578 in / 14630 out tokens · 152371 ms · 2026-08-04T20:40:12.359395+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of VRAE: Vertical Residual Autoencoder for License Plate Denoising and Deblurring." pith.science (2026). https://pith.science/paper/7FXYNBXE

@misc{pith2026250908392,
  author       = {Pith},
  title        = {Pith review of: VRAE: Vertical Residual Autoencoder for License Plate Denoising and Deblurring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7FXYNBXE}},
  note         = {Machine review of arXiv:2509.08392}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In real-world traffic surveillance, vehicle images captured under adverse weather, poor lighting, or high-speed motion often suffer from severe noise and blur. Such degradations significantly reduce the accuracy of license plate recognition systems, especially when the plate occupies only a small region within the full vehicle image. Restoring these degraded images a fast realtime manner is thus a crucial pre-processing step to enhance recognition performance. In this work, we propose a Vertical Residual Autoencoder (VRAE) architecture designed for the image enhancement task in traffic surveillance. The method incorporates an enhancement strategy that employs an auxiliary block, which injects input-aware features at each encoding stage to guide the representation learning process, enabling better general information preservation throughout the network compared to conventional autoencoders. Experiments on a vehicle image dataset with visible license plates demonstrate that our method consistently outperforms Autoencoder (AE), Generative Adversarial Network (GAN), and Flow-Based (FB) approaches. Compared with AE at the same depth, it improves PSNR by about 20%, reduces NMSE by around 50%, and enhances SSIM by 1%, while requiring only a marginal increase of roughly 1% in parameters.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 26 canonical work pages · 1 internal anchor

  1. [1]

    In: Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Pan, J., Sun, D., Pfister, H., Yang, M.-H.: Blind image deblurring using dark channel prior. In: Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1628–1636. IEEE (2016)

  2. [2]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pp

    Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pp. 8183–8192. IEEE (2018)

  3. [3]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Nah, S., Kim, T.H., Lee, K.M.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 257–265. IEEE (2017)

  4. [4]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018). CVPR 2018

  5. [5]

    EnlightenGAN: Deep Light Enhancement without Paired Supervision

    Jiang, Y., Gong, X., Liu, D., Cheng, Y., Fang, C., Shen, X., Yang, J., Zhou, P.: Enlightengan: Deep light enhancement without paired supervision. arXiv preprint arXiv:1906.06972 (2019). arXiv 2019

  6. [6]

    In: European Conference on Computer Vision (ECCV) (2020)

    Zamir, S.W., Arora, A., Long, C., Sohail, A., Khan, S., Bhulla, M., Khan, F.S., Yang, M.-H., Khan, M.A.: Learning enriched features for real image restoration and enhancement. In: European Conference on Computer Vision (ECCV) (2020). ECCV 2020

  7. [7]

    IEEE Transactions on Image Processing26(7), 3142–3155 (2017)

    Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing26(7), 3142–3155 (2017)

  8. [8]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops (2020)

    Anwar, S., Barnes, N.: Real image denoising with feature attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops (2020). ICCV Workshops 2020

  9. [9]

    arXiv preprint arXiv:1804.02767 (2018)

    Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018). arXiv 2018

  10. [10]

    In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2017)

    He, K., Gkioxari, G., Doll´ ar, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2017). ICCV 2017

  11. [11]

    Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.-A.: Extracting and composing robust features with denoising autoencoders, 1096–1103 (2008). ACM 10

  12. [12]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 8878–8887 (2019)

  13. [13]

    In: Advances in Neural Information Processing Systems, pp

    Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. In: Advances in Neural Information Processing Systems, pp. 10215–10224 (2018)

  14. [14]

    International Conference on Learning Representations (ICLR) (2025)

    Martin, S., Gagneux, A., Hagemann, P., Steidl, G.: Pnp-flow: Plug-and-play image restoration with flow matching. International Conference on Learning Representations (ICLR) (2025)

  15. [15]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Zhu, Y., Zhao, W., Li, A., Tang, Y., Zhou, J., Lu, J.: Flowie: Efficient image enhancement via rectified flow. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 13–22 (2024)

  16. [16]

    In: Proceedings of the 25th International Conference on Machine Learning, pp

    Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.-A.: Extracting and composing robust fea- tures with denoising autoencoders. In: Proceedings of the 25th International Conference on Machine Learning, pp. 1096–1103 (2008). ACM

  17. [17]

    In: Proceedings of the 28th ACM International Conference on Multimedia, pp

    Xu, J., Li, Z., Guo, H., He, H.: Lightweight image super-resolution with information multi-distillation network. In: Proceedings of the 28th ACM International Conference on Multimedia, pp. 1237–1245 (2020)

  18. [18]

    Remote Sensing10(11), 1700 (2018)

    Jiang, K., Wang, Z., Yi, P., Wang, G., Lu, T., Jiang, J.: Deep distillation recursive network for remote sensing image super-resolution. Remote Sensing10(11), 1700 (2018)

  19. [19]

    IEEE Transactions on Pattern Analysis and Machine Intelligence35(8), 1798–1828 (2013)

    Bengio, Y., Courville, A., Vincent, P.: Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence35(8), 1798–1828 (2013)

  20. [20]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Zhang, K., Zuo, W., Zhang, L.: Learning a single convolutional super-resolution network for multiple degradations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3262–3271 (2018)

  21. [21]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Ledig, C., Theis, L., Husz´ ar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., Shi, W.: Photo-realistic single image super-resolution using a generative adversarial network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4681–4690 (2017)

  22. [22]

    Proceedings of the ECCV Workshops (2018)

    Wang, X., Yu, K., Dong, C., Loy, C.C.: Esrgan: Enhanced super-resolution generative adversarial networks. Proceedings of the ECCV Workshops (2018)

  23. [23]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Zhang, K., Van Gool, L., Timofte, R.: Invertible image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10085–10094 (2021)

  24. [24]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778 (2016)

  25. [25]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4700–4708 (2017)

  26. [26]

    European Conference on Computer Vision (ECCV) Workshops, 255–271 (2018)

    Xu, Z., Yang, W., Meng, L., Lu, W., Huang, J., Zhou, J.: Towards end-to-end license plate detection and recognition: A large dataset and baseline. European Conference on Computer Vision (ECCV) Workshops, 255–271 (2018)

  27. [27]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Isola, P., Zhu, J.-Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversar- ial networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1125–1134 (2017)

  28. [28]

    In: ICLR (2017)

    Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real nvp. In: ICLR (2017)

  29. [29]

    Information Sciences681, 121239 (2024) https://doi.org/10.1016/j.ins.2024.121239 11

    Meni, M.J., White, R.T., Mayo, M.L., Pilkiewicz, K.R.: Entropy-based guidance of deep neural net- works for accelerated convergence and improved performance. Information Sciences681, 121239 (2024) https://doi.org/10.1016/j.ins.2024.121239 11