Pith. sign in

REVIEW 1 major objections 23 references

A Comparison of Data Augmentation Methods for Training Deep Neural Networks on Synthetic Aperture Sonar

T0 review · 1 major / 0 minor · reviewed 2026-07-30 · grok-4.5

Pith's one-line read Tuned data augmentations raise SAS target recognition accuracy, but gains depend on the network and do not always transfer.

desk verdict Useful SAS augmentation bake-off with a real SWIN lift; the ResNet comparison is too weak to carry the architecture claims. read the letter →

arxiv 2607.23770 v3 pith:BD2OOI54 submitted 2026-07-26 cs.LG

classification cs.LG
keywords syntheticaperturesonarautomatictargetrecognitiondataaugmentationdeepneuralnetworksSWINtransformerResNetphysics-informedMixupCutMixResize
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

Deep networks for automatic target recognition in synthetic aperture sonar are starved of labeled target examples because real SAS collection is expensive. This paper systematically compares twelve augmentations—physics-motivated flips and noise models plus modern computer-vision mixes such as Mixup, CutMix, and ResizeMix—after tuning each one's hyperparameters on a SWIN transformer. Nearly every single augmentation helps the transformer; a short greedy policy (Zoom, then ResizeMix, small rotation, contrast shift) lifts its average precision from about 75 to 86. The same policy and some individual transforms fail or even hurt a ResNet of similar size, while the transformer substantially outperforms the convolutional baseline under the shared training recipe. The work therefore shows that augmentation still matters for large modern models on SAS, but the benefit is architecture-dependent and policies found on one network need not transfer.

What carries the argument

Greedy sequential forward search that builds a multi-augmentation policy by repeatedly adding the single remaining transform that most improves validation performance, starting from a no-augmentation baseline and using hyperparameters first tuned on SWIN-base.

What would settle it

Re-optimize every augmentation hyperparameter and re-run the greedy policy search on ResNet-152 alone; if the resulting ResNet policy then matches or exceeds the SWIN gains, the claimed architecture–augmentation interaction is an artifact of transfer rather than a true architectural difference.

Watch

Extended reading notes

Core claim

With hyperparameters optimized on SWIN-base, nearly all twelve tested augmentations improve AP or F1 for SAS-ATR on that transformer; a greedy multi-augmentation policy further raises SWIN AP from 74.82 (no augmentation) to 86.33, yet the identical policy yields no gain on ResNet-152 and some single augmentations (Mixup, Speckle) harm it, while SWIN itself substantially outperforms ResNet under the same training setup.

Load-bearing premise

Hyperparameters and the multi-augmentation policy tuned only on the SWIN transformer can be transferred unchanged to ResNet-152 when claiming that architecture controls whether an augmentation helps.

Editorial extensions

If this is right

  • Practitioners can safely start with Zoom, ResizeMix, small rotation and contrast shift when training transformer SAS-ATR models.
  • Physics fidelity is not required for an augmentation to help: several pure computer-vision mixes improve accuracy.
  • Augmentation policies found on one modern architecture should be re-validated before being applied to another.
  • Transformer backbones merit wider use in SAS-ATR literature given their large margin over a comparable ResNet under identical training.

Reading between the lines

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

  • The large unaugmented performance gap (SWIN AP 74.8 vs ResNet 29.6) suggests the shared ImageNet-pretrained recipe may already favor transformers, so architecture comparisons on SAS may need architecture-specific optimizers or schedules.
  • Because the greedy search is known to be suboptimal, an exhaustive or evolutionary search over the same twelve transforms could yield still stronger SAS policies.
  • If re-tuning on ResNet restores the gains, the practical takeaway becomes “always tune augmentations per architecture” rather than “transformers uniquely benefit.”
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

1 major / 0 minor

Summary. The manuscript compares twelve computer-vision and physics-informed augmentations for binary SAS automatic target recognition on 37,271 prescreener snippets. Using ImageNet-pretrained SWIN-base and ResNet-152 models, it tunes a small hyperparameter grid for each augmentation on SWIN, evaluates individual augmentations, and constructs a four-augmentation policy by greedy forward selection. It reports that most augmentations improve SWIN, that the selected policy raises SWIN AP from 74.82 to 86.33, that the SWIN-selected settings do not improve ResNet-152, and that SWIN substantially outperforms ResNet. The authors conclude that many augmentations are useful despite imperfect physical fidelity, that their benefits can be partially additive, and that augmentation effects are architecture-dependent.

Significance. If the experimental protocol is validated, this is a useful empirical benchmark for a data-limited SAS-ATR setting. Its strengths are the broad side-by-side comparison of twelve augmentations, explicit per-augmentation hyperparameter tuning rather than untuned defaults, inclusion of recent mixing augmentations alongside SAS-motivated transformations, and an explicit greedy search over augmentation policies. The use of a modern pretrained transformer and the direct test of cross-architecture transfer are also relevant to a literature dominated by CNNs. The paper is candid that greedy search is suboptimal and that ResNet-specific tuning might change transfer results. At present, however, the strongest support is for SWIN-specific findings; the architecture and transfer conclusions remain vulnerable to the comparator and evaluation concerns below.

major comments (1)
  1. [§3 Training, Table 5, and Conclusion] Both architectures use one fixed recipe—Adam-W, batch size 64, warmup from 3e-5 to 3e-4, cosine annealing, and 100 epochs—while all augmentation hyperparameters are selected on SWIN. The resulting ResNet baseline of AP 29.57/F1 52.00 is sufficiently anomalous that an undertrained or recipe-mismatched comparator is plausible. Consequently, Table 5’s transfer failure, the reported Mixup/Speckle harm, and the conclusion that SWIN outperforms ResNet “by substantial margins” cannot yet be attributed to architecture. Please retrain ResNet with an established or independently tuned recipe, report optimization/validation curves, and ideally tune the augmentation settings on ResNet, or narrow these conclusions accordingly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical augmentation bake-off on held-out folds; metrics are external, not defined from fitted hyperparameters.

full rationale

This paper is a systematic empirical comparison of CV and physics-informed data augmentations for SAS-ATR on SWIN-base and ResNet-152. The load-bearing claims are measured AP and F1 on randomized two-fold cross-validation of a fixed labeled dataset, not quantities derived from first-principles equations. Hyperparameter grids (Table 3) and the greedy multi-augmentation policy (Table 4) are selected on SWIN and then scored with the same external metrics; that is ordinary validation practice, not a fitted input renamed as a prediction, and no equation reduces a reported score to a tuned constant by construction. Prior SAS/CV citations supply candidate transforms and baselines, not uniqueness theorems or self-justifying premises that force the results. Architecture comparison and transfer failure are likewise empirical observations under a shared training recipe. There is no self-definitional loop, no ansatz smuggled in as derivation, and no renaming of a known identity as a novel prediction. Circularity score is therefore 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

Load-bearing content is experimental protocol, not new theory. The claim rests on standard ML training assumptions, a private SAS dataset, ImageNet-pretrained backbones, a coarse hyperparameter grid chosen on one model, and a greedy policy search. No new physical entities. Free parameters are the usual training and augmentation knobs that were selected by validation performance on SWIN.

free parameters (4)
  • Per-augmentation hyperparameter grids (scale, translate, degrees, contrast_factor, brightness/contrast, filterSize/rate, = Chosen values bolded in Table 3 (e.g. Zoom scale min/max 0.8/1.2-class settings as selected; alphas etc.)
    Three-to-four candidate values per hyperparameter; the value with highest SWIN performance is locked for all later experiments (Table 3).
  • Learning-rate schedule and optimization hyperparameters = 3e-5 to 3e-4 warmup; AdamW; focal loss
    Warmup 3e-5→3e-4 over 10 epochs, cosine anneal, AdamW, batch 64, 100 epochs, focal loss; fixed by authors, not swept in the paper.
  • Greedy multi-augmentation policy order and stopping = Zoom + ResizeMix + small Rotation + Contrast Shift (AP 86.33)
    Sequential forward selection on SWIN until AP stops improving; four-step policy retained as the reported multi-aug result.
  • Preprocessing blur / resize / normalization = 256×256 magnitude snippets
    Small Gaussian denoising blur, min-max norm, bilinear resize to 256×256, magnitude-only input—fixed pipeline choices that affect all runs.
assumptions (5)
  • domain assumption Two-fold randomized cross-validation on this prescreened snippet set is a sufficient estimate of generalization for ranking augmentations.
    Sec. 3 Dataset and Preprocessing; only two folds, same 1:24 class ratio, no external test set.
  • domain assumption ImageNet-1k pretrained SWIN-base and ResNet-152 with the shared MMPretrain recipe are fair, representative modern backbones for SAS-ATR comparison.
    Sec. 2.2; capacity is similar (~60M vs ~88M) but domain shift from optical pretraining is unablated.
  • domain assumption Using only SAS magnitude (discarding phase/complex structure) is adequate for the ATR task under study.
    Sec. 3 preprocessing: ‘we only use the magnitude data.’
  • ad hoc to paper Physics-informed vs CV labels for augmentations follow the paper’s asterisk convention and prior SAS/SAR citations; beneficial need not imply physical fidelity.
    Table 1 and Conclusion argue physical consistency is not required for gains—an interpretive stance of this paper.
  • standard math Standard supervised learning bounds / empirical risk ideas justify augmentation as i.i.d.-style expansion of scarce target chips.
    Implicit throughout Introduction and Methods; no new learning-theoretic derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Comparison of Data Augmentation Methods for Training Deep Neural Networks on Synthetic Aperture Sonar." pith.science (2026). https://pith.science/paper/BD2OOI54

@misc{pith2026260723770,
  author       = {Pith},
  title        = {Pith review of: A Comparison of Data Augmentation Methods for Training Deep Neural Networks on Synthetic Aperture Sonar},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BD2OOI54}},
  note         = {Machine review of arXiv:2607.23770}
}
read the original abstract

In this work we study Automatic Target Recognition (ATR) for Synthetic Aperture Sonar (SAS) data with a focus on deep neural networks (DNNs). The main challenge in training DNNs for SAS-ATR arises from the limited quantity of labeled target examples due to the significant costs and time required to collect real-world SAS data. One successful general strategy for mitigating the problem of limited training data is augmentation, which generates additional synthetic training data by introducing realistic variations to available data. Prior research has investigated a variety of augmentation strategies for SAS-ATR, including conventional image augmentations (e.g., contrast changes, cropping) as well as augmentations motivated the specific physics of SAS data. Building on prior work, we systematically compare many of these existing augmentation strategies for training DNNs for SAS-ATR. We also investigate the impact of augmentation when combined with modern DNN architectures such as transformers. The results indicate that augmentation can improve target recognition accuracy, although benefits vary, and not all augmentations are beneficial.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 1 linked inside Pith

  1. [1]

    Spectral Partitioning of Synthetic Aperture Sonar Imagery for Improved ATR

    D.P. Williams and D.C. Brown, “Spectral Partitioning of Synthetic Aperture Sonar Imagery for Improved ATR”, IEEE Geoscience and Remote Sensing Letters, Vol. 22, (2025)

  2. [2]

    Additional Representations for Improving Synthetic Aperture Sonar Classification Using Convolutional Neural Networks

    I.D. Gerg and D.P. Williams, “Additional Representations for Improving Synthetic Aperture Sonar Classification Using Convolutional Neural Networks”, ICUA 2018, Vol. 40, pt 2. (2018)

  3. [3]

    Structural Prior Driven Regularized Deep Learning for Sonar Image Classification

    I.D. Gerg and V. Monga, “Structural Prior Driven Regularized Deep Learning for Sonar Image Classification”, IEEE Geoscience and Remote Sensing Letters, Vol. 60, (2022)

  4. [4]

    Target Detection and Segmentation in Circular-Scan Synthetic-Aperture- Sonar Images using Semi-Supervised Convolutional Encoder-Decoders

    I.J. Sledge et al., “Target Detection and Segmentation in Circular-Scan Synthetic-Aperture- Sonar Images using Semi-Supervised Convolutional Encoder-Decoders”, IEEE JOE, Vol. 47, (2022)

  5. [5]

    Supervised Deep Learning Classification for Multi -band Synthetic Aperture Sonar

    M.S. Emigh et al., “Supervised Deep Learning Classification for Multi -band Synthetic Aperture Sonar”, ICUA, (2018)

  6. [6]

    Underwater Target Classification in Synthetic Aperture Sonar Imagery Using Deep Convolutional Neural Networks

    D.P. Williams, “Underwater Target Classification in Synthetic Aperture Sonar Imagery Using Deep Convolutional Neural Networks”, ICPR, (2016)

  7. [7]

    On the Use of Tiny Convolutional Neural Networks for Human -Expert-Level Classification Performance in Sonar Imagery

    D.P. Williams, “On the Use of Tiny Convolutional Neural Networks for Human -Expert-Level Classification Performance in Sonar Imagery”, IEEE JOE, Vol. 46, NO.1. (2021)

  8. [8]

    A survey on image data augmentation for deep learning

    C. Shorten and T. M. Khoshgoftaar, “A survey on image data augmentation for deep learning” Big Data, Vol 6, no. 1, pp.1-48, (2019)

Show all 23 references
  1. [9]

    Convolutional Neural Network with Data Augmentation for SAR Target Recognition

    J. Ding et a l., “Convolutional Neural Network with Data Augmentation for SAR Target Recognition”, IEEE Geoscience and Remote Sensing Letters, Vol. 13, (2016)

  2. [10]

    Improving SAS Automatic Object Recognition With Sub-Aperture Imagery From Circular SAS

    M.S. Emigh et al., “Improving SAS Automatic Object Recognition With Sub-Aperture Imagery From Circular SAS”, ICUA, (2024)

  3. [11]

    A Study on the Effect of Commonly Used Data Augmentation Techniques on Sonar Image Artifact Detection Using Deep Neural Network s

    M. Oresca nin et al. , “A Study on the Effect of Commonly Used Data Augmentation Techniques on Sonar Image Artifact Detection Using Deep Neural Network s”, IEEE International Geoscience and Remove Sensing Symposium, (2023)

  4. [12]

    Mixup: Beyond Empirical Risk Minimization

    H Zhang et al., “Mixup: Beyond Empirical Risk Minimization”, ICLR, (2018)

  5. [13]

    ResizeMix: Mixing Data with Preserved Object Information and True Labels

    J. Qin et al. , “ResizeMix: Mixing Data with Preserved Object Information and True Labels” , arXiv arxiv.org/abs/2012.11101, (2020)

  6. [14]

    Yun et al ., “CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features, ICCV, (2019)

    S. Yun et al ., “CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features, ICCV, (2019)

  7. [15]

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows

    Z. Liu e t al., “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows” , IEEE ICCV, (2021)

  8. [16]

    Deep Residual Learning for Image Recognition

    K. He et al., “Deep Residual Learning for Image Recognition”, IEEE CVPR, (2016)

  9. [17]

    OpenMMLab’s Pre -training Toolbox and Benchmark

    MMPretrain Contributors , “ OpenMMLab’s Pre -training Toolbox and Benchmark ”, GitHub, https://github.com/open-mmlab/mmpretrain, (2023)

  10. [18]

    TorchVision: Pytorch computer vision library

    PyTorch Contributors , “ TorchVision: Pytorch computer vision library ”, GitHub, https://github.com/pytorch/vision

  11. [19]

    A Speckle Reduction Method Based on Coupled Partial Differential Equation for Synthetic Aperture Sonar

    P. Huang and H. Zhong, “A Speckle Reduction Method Based on Coupled Partial Differential Equation for Synthetic Aperture Sonar”, CISP-BMEI, (2019)

  12. [20]

    ImageNet: A Large-Scale Hierarchical Image Database

    J. Deng et al., “ImageNet: A Large-Scale Hierarchical Image Database”, IEEE CVPR, (2009)

  13. [21]

    An Introduction to Information Retrieval

    C. Manning et al. , “An Introduction to Information Retrieval ”, Cambridge University Press , (2009)

  14. [22]

    Decoupled Weight Decay Regularization

    I. Loshchilov and F. Hutter, “Decoupled Weight Decay Regularization”, ICLR, (2019)

  15. [23]

    Focal Loss for Dense Object Detection

    T. Lin et al., “Focal Loss for Dense Object Detection”, IEEE ICCV, (2017)

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.