Pith. sign in

REVIEW 7 major objections 5 minor 59 references

FORTRESS: Function-composition Optimized Real-Time Resilient Structural Segmentation via Kolmogorov-Arnold Enhanced Spatial Attention Networks

T0 review · 7 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 FORTRESS claims state-of-the-art defect segmentation (F1 0.771) at 91% lower cost than U-Net, by replacing standard convolutions with depthwise separable ones and adding selective Kolmogorov-Arnold transformations at bottleneck layers.

desk verdict A plausible lightweight segmentation architecture with real efficiency gains, but the state-of-the-art claims don't survive the baseline tables; needs major revision, not desk rejection. read the letter →

arxiv 2507.12675 v1 pith:4MHLJISC submitted 2025-07-16 cs.CV cs.AIeess.IV

Christina Thrainer , Md Meftahul Ferdaus , Mahdi Abdelguerfi , Christian Guetl , Steven Sloan , Kendall N. Niles , Ken Pathak This is my paper
classification cs.CVcs.AIeess.IV
keywords structuraldefectsegmentationKolmogorov-Arnoldnetworksdepthwiseseparableconvolutionspatialattentionparameterefficiencyreal-timesewerdatasetU-Netencoder-decoder
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

The paper tries to show that pixel-level structural defect segmentation does not have to choose between accuracy and real-time speed. It presents FORTRESS, a U-Net-style network whose standard convolutions are replaced by depthwise separable ones, and which inserts tiny Kolmogorov-Arnold transformations only at bottleneck layers where they are cheap. On the CSDD sewer-defect benchmark the network reports an F1 score of 0.771 without background and a mean IoU of 0.677 with background, both above SA-UNet, its strongest competitor, while using 2.9M parameters and 1.17 GFLOPs, roughly 91% less than plain U-Net. The authors argue the same design stays accurate with only a quarter of the training data and transfers to a second concrete-defect dataset, which matters because real infrastructure inspection needs models small enough to run on drones and edge devices.

What carries the argument

The load-bearing mechanism is the adaptive TiKAN gating condition: the Kolmogorov-Arnold module activates only when the feature map has at least 16 channels and a spatial area no greater than 1024 pixels, and is skipped otherwise, so spline-based function composition is confined to low-resolution, high-channel stages. TiKAN itself is a Tiny Kolmogorov-Arnold Network, meaning learnable univariate spline functions with low-rank factorization, combined with a depthwise enhance convolution. The second pillar is the depthwise separable convolution, whose parameter cost is $C_{in}\cdot 9 + C_{in}\cdot C_{out}$ instead of $C_{in}\cdot C_{out}\cdot 9$, giving a per-layer reduction factor of $9C_{out}/(9+C_{out})$ that approaches 9 when $C_{out}$ is large. Multi-scale attention fusion then concatenates spatial attention, channel attention, and TiKAN-enhanced features at each decoder level and blends them with a learned 1x1 convolution, with deep supervision weights 0.4, 0.3, and 0.2 stabilizing training.

What would settle it

Train FORTRESS and SA-UNet on CSDD under identical per-model hyperparameter tuning, including the same Dynamic Label Injection budget, and compare F1 and mIoU on the official test split; if SA-UNet, U-KAN, or U-Net reaches or exceeds 0.771 F1 without background and 0.677 mIoU with background, the claimed state-of-the-art advantage disappears. A quicker check is to rerun the no-augmentation protocol of Table V with all baselines using their own recommended optimizers and learning rates, since the near-zero scores there (SegFormer 0.019, MobileUNETR 0.017, FasterViT 0.030) should recover to plausible levels under fair tuning.

Watch

Extended reading notes

Core claim

On its own terms, FORTRESS's central discovery is that the accuracy/efficiency trade-off in structural defect segmentation can be broken by a dual optimization: systematic depthwise separable convolutions cut parameters by roughly 3.6 times per convolutional layer, while an adaptive TiKAN module adds Kolmogorov-Arnold function-composition capacity only when the feature map is small enough (area at most 1024 pixels) and deep enough (16 or more channels) for that computation to be cheap. The resulting 2.89M-parameter network reports F1 of 0.771 without background and mIoU of 0.677 with background on CSDD, above SA-UNet and U-KAN, and 0.778 F1 on S2DS, above BiFPN, while using 1.17 and 4.37 GFLOPs respectively. The paper additionally reports that the network keeps the lead when training data is cut to 50% or 25% and when augmentation is removed, which it reads as evidence that the efficiency comes from the architecture rather than from the data pipeline.

Load-bearing premise

The state-of-the-art margin over baselines assumes every comparison model was trained with comparably favorable settings, seeds, and augmentation; several baselines collapse to near-zero F1 in the paper's data-limited and no-augmentation tables, which points to training instability or mismatched hyperparameters rather than architectural inferiority.

Editorial extensions

If this is right

  • Deploying FORTRESS replaces a 31M-parameter, 13.69-GFLOP U-Net with a 2.89M-parameter, 1.17-GFLOP model that reports higher F1 and mIoU on CSDD, so real-time inspection hardware can run a more accurate segmenter.
  • At 25% of the CSDD training data, FORTRESS reports F1 of 0.669 without background, roughly 3.6 points above SA-UNet, implying the architecture is data-efficient enough for inspection regimes with little labeled data.
  • On the S2DS concrete-defect set, FORTRESS reports F1 of 0.778 without background and 94% fewer GFLOPs than BiFPN, indicating the design generalizes across infrastructure types rather than overfitting sewer imagery.
  • The adaptive TiKAN gate confines function-composition computation to high-channel, low-resolution feature maps, which is the stated reason the network gets KAN-style representational power without KAN's usual cost.

Reading between the lines

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

  • The gating rule $C_{in} \ge 16$ and $H\times W \le 1024$ is a cost-geometry heuristic: it turns KAN processing on exactly where spatial size is small enough to keep spline evaluation cheap and channel count high enough to amortize the transformation. The same rule could be lifted into any encoder-decoder, suggesting a general recipe for adding KAN blocks to lightweight segmentation without trainin
  • Because the reported speed (110 FPS on GPU, 9.2 FPS on CPU) is measured at 256x256 input, real deployments at higher inspection resolutions would push TiKAN past its activation threshold and lose the KAN contribution; testing at native resolutions would show whether the accuracy advantage persists where the mechanism is off.
  • The near-zero baseline scores in the no-augmentation and 25%-data tables suggest the headline margin may owe as much to training stability under Dynamic Label Injection as to the architecture itself; a fair-tuning replication could separate those two contributions.
  • If the data-efficiency result holds, FORTRESS's size makes it a natural candidate for on-device federated inspection, letting drones share model updates without uploading raw defect imagery.
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

7 major / 5 minor

Summary. The paper proposes FORTRESS, a U-Net-like encoder-decoder for structural defect segmentation that combines depthwise separable convolutions, an adaptively gated 'TiKAN' Kolmogorov-Arnold module, multi-scale spatial/channel attention, and deep supervision. The authors report large efficiency gains (2.89M parameters, 1.17 GFLOPs, roughly 3x faster inference than a conventional U-Net) and claim state-of-the-art segmentation accuracy, with an F1-score of 0.771 and mIoU of 0.677 on the CSDD benchmark, cross-dataset results on S2DS, and robustness ablations at reduced training-data fractions. The paper includes detailed hyperparameters, hardware requirements, and a source-code URL.

Significance. If the reported results survive scrutiny, FORTRESS would be a practically useful contribution: a roughly 91% parameter/FLOP reduction relative to a standard U-Net at near-identical or better segmentation quality is valuable for real-time infrastructure inspection. The authors also provide unusually detailed implementation specifications, which supports reproducibility. However, the empirical claims of state-of-the-art performance currently rest on baseline comparisons that are internally inconsistent, and several central equations are mutually contradictory, so the significance is conditional on a substantive re-evaluation rather than on cosmetic revision.

major comments (7)
  1. [Abstract and Table I] The abstract's headline numbers mix incompatible metrics: it reports 'F1-score of 0.771 and a mean IoU of 0.677,' but Table I lists F1 = 0.771 only when background is excluded and mIoU = 0.677 only when background is included; the mIoU without background is 0.643. Because the paper's central SOTA claim is stated in terms of these two numbers, the metric definitions must be made consistent everywhere, including the abstract and conclusion.
  2. [Tables I and II] The same baseline names have different parameter counts on the two datasets: Swin-UNet is 14.50M on CSDD but 2.63M on S2DS, SegFormer is 13.67M vs 2.67M, EGE-UNet is 3.02M vs 2.83M, and FasterViT is 25.23M vs 23.83M. Parameter counts are inherent to a model architecture; they cannot vary with the dataset unless different model variants were used. This means the cross-dataset comparison in Table II is not a comparison of the same models, and it undermines the general 'state-of-the-art' claim across benchmarks.
  3. [Tables IV and V] Several baselines collapse to near-zero F1 scores under the robustness conditions (Rolling UNet-L: 0.112 in Table IV and 0.121 in Table V; MobileUNETR: 0.133 and 0.119; SegFormer: 0.121 in Table IV; FasterViT: 0.127 and 0.131), while FORTRESS remains at 0.669/0.759. Such abrupt collapses are characteristic of broken training configurations or incompatible hyperparameters, not of the gradual degradation one would expect at 25% data or without augmentation. Because FORTRESS's margin over the best functioning baseline (SA-UNet) is only 0.7-1.0 points in the primary tables, a fair tuning of these baselines could erase the claimed advantage; the paper must demonstrate a fixed, standardized protocol for all methods.
  4. [Section IV-C, Eqs. (20)-(22)] Equations (20) and (21) pool over the spatial dimensions H and W to produce tensors of shape 1x1xC, but Eq. (22) then states that the concatenation of these pooled features is in R^{HxWx2}. This is mathematically inconsistent: a feature map pooled over H and W no longer has spatial extent HxW. The spatial-attention formulation, which is one of the paper's central contributions, needs to be rewritten with consistent tensor shapes, e.g., pooling over channels to obtain HxWx1 maps before concatenating into an HxWx2 map.
  5. [Eqs. (1) and (10)] The 'dual optimization strategy' is introduced in Eq. (1) as minimizing L_seg(phi) + lambda_eff * C_comp(phi), where C_comp is a computational-complexity penalty. However, the actual training loss in Eq. (10) (and again in Eq. (25)) contains only cross-entropy terms and deep-supervision weights; no C_comp term appears, and the value of lambda_eff is never specified or used. The paper needs to either implement this term in the training objective or clearly explain that Eq. (1) is only a conceptual framing rather than the optimized objective.
  6. [Eqs. (3), (12) and Section IV-A] The reduction-factor formula in Eqs. (3) and (12) gives 9*C_out/(9+C_out), which approaches 9 for large C_out, yet the abstract and Section I claim a '3.6x parameter reduction per layer.' The paper also states in Section IV-A that the design gives 'approximately 65% parameter reduction compared to conventional attention-based U-Net variants.' These numbers are mutually inconsistent and need a single, clearly derived accounting of the parameter reduction, including whether the 3.6x figure is per layer or network-wide.
  7. [Section VI-E] The extended cross-dataset results on bridge, tunnel, and building-facade imagery report specific F1 and mIoU values (0.742/0.618, 0.728/0.605, 0.751/0.629) but provide no dataset names, sizes, splits, or result tables. In addition, the claims of 'McNemar's test p < 0.001' and 'bootstrap confidence intervals ranging from 0.8 to 4.2%' in Section VI-E are given without any methodological detail, and Section V-B's promise of 'multiple random seeds' is not backed by reported variances or seed values. Statistical significance claims cannot be verified without this information.
minor comments (5)
  1. [Abstract] There is a typo in 'F1- score' (the hyphen is separated from 'F1').
  2. [Section VII-A] The hyperparameter section introduces an 'adaptive weighting decay constant tau = 1000' with no corresponding formula or description of how it affects the deep-supervision weights; please define its role.
  3. [Section VII-B] The hardware section reports that CSDD training for 50 epochs completes in 0.1 hours, but at 3.99 batches/second with roughly 194 batches per epoch, 50 epochs would take about 0.68 hours; this inconsistency should be corrected.
  4. [Section V-A] The description of Dynamic Label Injection states that CSDD has 'no purely defect-free images' and that 155 background-only samples were cropped, expanding the training set to 3,106 images; the arithmetic and the relationship between 580 videos, 12,230 images, and the 70/15/15 split should be clarified.
  5. [Section VI-C] In the training-dynamics discussion, the text says the validation IoU reaches 'approximately 0.77' on S2DS, but Table II lists FORTRESS's mIoU on S2DS as 0.697 (with background); please clarify whether Figure 4 reports a different IoU variant.

Circularity Check

1 steps flagged · score 2.0 of 10

No material circularity: FORTRESS's efficiency and accuracy numbers are measured outputs, not derivations reduced to their inputs; the one self-citation (the authors' own CSDD benchmark) is not load-bearing because the external S2DS benchmark independently supports the central claim, while baseline-fairness and missing-ablation issues are correctness risks, not circularity.

  1. self citation load bearing [Section V.A (Datasets); Abstract; References [27]]
    "The primary dataset used for training and comprehensive evaluation is the Culvert Sewer Defect Dataset (CSDD) [27], which was developed for autonomous defect detection in real-world infrastructure."

    Ref [27] (Alshawi et al., 'Dual attention u-net with feature infusion') is the CSDD source and its authors overlap with this paper (Ferdaus, Abdelguerfi, Niles, Prathak); the Abstract's headline F1 0.771 / mIoU 0.677 are measured on that self-authored benchmark, so the central SOTA claim leans on the authors' own dataset. Minor rather than load-bearing: the dataset is a fixed annotated artifact, the metrics are empirical test-split measurements with no definitional tie to FORTRESS, and the method independently tops the external S2DS benchmark (Table II: 0.778 F1 w/o bg vs BiFPN 0.771), so the central claim has independent content beyond the self-citation.

full rationale

FORTRESS's load-bearing results are measured outputs, not equations that equal their own inputs. The efficiency claims trace to Eq. (3)/(12), which is the standard depthwise-separable-versus-standard convolution parameter ratio, 9*Cout/(9+Cout); the 3.6x per-layer figure and the 91% totals are arithmetic on measured parameter and FLOP counts, so they do not presuppose the claimed outcome. The adaptive TiKAN gate (Eq. 6, Eq. 19; thresholds gamma_c=16, gamma_s=1024) is a hand-set architectural criterion rather than a fitted parameter later renamed a prediction, and no quantity is fit to a subset and then 'predicted' on a closely related one. The Kolmogorov-Arnold premise is attributed to the external representation theorem (refs [33][34]) and to Liu et al. [35]; no author-uniqueness theorem is invoked, and no ansatz is smuggled in via self-citation. The one genuine self-citation is the CSDD benchmark (ref [27], overlapping authors), used for the headline numbers; this is scored as minor because the dataset is a fixed annotated artifact, the metrics are measured on an official held-out split, and the same architecture independently ranks first on the external S2DS benchmark, so the central claim has independent content. Concerns that do not qualify as circularity should be weighed separately as correctness risks. First, the section titled 'Ablation Study' (VI.D, Tables III-V) contains data-efficiency and no-augmentation robustness runs, not component ablations of TiKAN, depthwise separability, or attention fusion; the statement that 'the dual optimization strategy proves essential' (Abstract/Conclusion) is therefore asserted without the isolating evidence the section name promises. Second, baseline fairness is questionable from the paper's own tables: Swin-UNet is listed with 14.50M parameters (Table I) but 2.63M (Table II), SegFormer with 13.67M vs 2.67M, U-Net with 31.04M vs 26.08M, and several baselines collapse to 0.112-0.133 F1 in Tables IV-V, which is characteristic of protocol mismatch rather than gradual degradation. Third, the Abstract mixes metrics, pairing F1 without background (0.771) with mIoU including background (0.677). Since the CSDD margin over SA-UNet is only 0.7-1.0 points, these risks could erode the SOTA claim, but none of them exhibits a result reducing by construction to its inputs, so the circularity score remains low.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central efficiency gains come from standard depthwise separable convolutions, which require no new assumptions. The KAN-related components introduce hand-chosen thresholds and an under-defined module. The training objective in Eq (1) is not used, and the comparative evaluation relies on baseline training assumptions that appear violated.

free parameters (5)
  • TiKAN gating thresholds gamma_c and gamma_s = 16 channels, 1024 pixels
    Hand-chosen in Eq (19) to decide when KAN is active; no sensitivity analysis is provided, and the threshold effectively deactivates KAN at most layers for 256x256 input.
  • Deep supervision weights beta_2, beta_3, beta_4 = 0.4, 0.3, 0.2
    Fixed constants in Eq (25), not tuned or ablated.
  • Class weights for CSDD cross-entropy = [1.0, 3.0, 1.0, 1.0, 1.2, 1.5, 3.0, 1.2, 1.3]
    Manually specified in Section VII-A, not based on a documented procedure.
  • TiKAN spline grid size G and order O = G=5, O=3
    TiKAN configuration given in Section VII-A without ablation.
  • Lambda_eff in the dual optimization objective (Eq 1) = not specified
    Eq (1) includes a complexity penalty with weight lambda_eff, but training uses only cross-entropy losses in Eq (25). The dual objective is never actually optimized, so the parameter is non-functional.
assumptions (4)
  • domain assumption Kolmogorov-Arnold representation theorem is a useful inductive bias for image segmentation
    Invoked in Section II-B and III as theoretical motivation, but no proof or controlled experiment shows that KAN improves defect segmentation over standard convolutions.
  • domain assumption The benchmark datasets and annotations are reliable and comparable across methods
    Section V uses CSDD (from the authors' group, ref [27]) and S2DS; comparability assumes the same splits and annotation quality.
  • domain assumption Baselines are re-implemented fairly with equivalent training
    Section V-B claims identical augmentation and protocols, but the collapse of several baselines to near-zero F1 in Tables IV and V suggests this assumption is violated.
  • domain assumption Input resolution of 256x256 and ImageNet normalization are appropriate for all architectures
    Section VII-B sets a single resolution and normalization for all models, which may disadvantage transformer-based baselines that often use larger resolution.
invented entities (2)
  • TiKAN module
    purpose: Kolmogorov-Arnold feature transformation applied conditionally to low-resolution features
    Introduced in this paper (Section II-B), no external validation, no ablation isolating its contribution to the reported performance.
  • Adaptive activation criterion (Cin >= 16 and HxW <= 1024)
    purpose: Gates when TiKAN is applied to features
    Hand-set thresholds in Eq (19); the paper provides no evidence that these thresholds are optimal or that KAN helps in the active regime.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FORTRESS: Function-composition Optimized Real-Time Resilient Structural Segmentation via Kolmogorov-Arnold Enhanced Spatial Attention Networks." pith.science (2026). https://pith.science/paper/4MHLJISC

@misc{pith2026250712675,
  author       = {Pith},
  title        = {Pith review of: FORTRESS: Function-composition Optimized Real-Time Resilient Structural Segmentation via Kolmogorov-Arnold Enhanced Spatial Attention Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4MHLJISC}},
  note         = {Machine review of arXiv:2507.12675}
}
read the original abstract

Automated structural defect segmentation in civil infrastructure faces a critical challenge: achieving high accuracy while maintaining computational efficiency for real-time deployment. This paper presents FORTRESS (Function-composition Optimized Real-Time Resilient Structural Segmentation), a new architecture that balances accuracy and speed by using a special method that combines depthwise separable convolutions with adaptive Kolmogorov-Arnold Network integration. FORTRESS incorporates three key innovations: a systematic depthwise separable convolution framework achieving a 3.6x parameter reduction per layer, adaptive TiKAN integration that selectively applies function composition transformations only when computationally beneficial, and multi-scale attention fusion combining spatial, channel, and KAN-enhanced features across decoder levels. The architecture achieves remarkable efficiency gains with 91% parameter reduction (31M to 2.9M), 91% computational complexity reduction (13.7 to 1.17 GFLOPs), and 3x inference speed improvement while delivering superior segmentation performance. Evaluation on benchmark infrastructure datasets demonstrates state-of-the-art results with an F1- score of 0.771 and a mean IoU of 0.677, significantly outperforming existing methods including U-Net, SA-UNet, and U- KAN. The dual optimization strategy proves essential for optimal performance, establishing FORTRESS as a robust solution for practical structural defect segmentation in resource-constrained environments where both accuracy and computational efficiency are paramount. Comprehensive architectural specifications are provided in the Supplemental Material. Source code is available at URL: https://github.com/faeyelab/fortress-paper-code.

Figures

Figures reproduced from arXiv: 2507.12675 by the authors.

Figure 1
Figure 1. FORTRESS architecture overview. The model employs a U-Net [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Comprehensive segmentation results on CSDD dataset showing [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 4
Figure 4. Training dynamics and convergence analysis for FORTRESS on [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 53 canonical work pages

  1. [1]

    Deep cnn-based visual defect detection: Survey of current literature,

    S. B. Jha and R. Babiceanu, “Deep cnn-based visual defect detection: Survey of current literature,” Computers in Industry, vol. 148, p. 103911, 2023

    work page 2023

  2. [2]

    Few-shot learning for structural health diagnosis of civil infrastructure,

    Y . Xu, Y . Fan, Y . Bao, and H. Li, “Few-shot learning for structural health diagnosis of civil infrastructure,” Advanced Engineering Informatics , vol. 62, p. 102650, 2024

    work page 2024

  3. [3]

    Defect detection in civil structure using deep learning method,

    X. Zhou and R. L. K. Tiong, “Defect detection in civil structure using deep learning method,” in Journal of Physics: Conference Series , vol. 2762, 2024

    work page 2024

  4. [4]

    Fusion of thermal and rgb images for automated deep learning based crack detection in civil infrastructure,

    Q. G. Alexander, V . Hoskere, Y . Narazaki, A. Maxwell, and B. Spencer, “Fusion of thermal and rgb images for automated deep learning based crack detection in civil infrastructure,” AI in Civil Engineering , vol. 1, 2022

    work page 2022

  5. [5]

    Learning monoc- ular depth estimation for defect measurement from civil rgb-d dataset,

    M. Midwinter, Z. Al-Sabbag, R. Bajaj, and C. Yeum, “Learning monoc- ular depth estimation for defect measurement from civil rgb-d dataset,” Structural Health Monitoring , 2025

    work page 2025

  6. [6]

    Deep learning-based concrete defects classification and detection using semantic segmentation,

    P. Arafin, A. M. Billah, and A. Issa, “Deep learning-based concrete defects classification and detection using semantic segmentation,” Struc- tural Health Monitoring , vol. 23, pp. 383 – 409, 2023

    work page 2023

  7. [7]

    Lightweight pixel-level semantic segmentation and analysis for sewer defects using deep learning,

    L. Dang, H. Wang, Y . Li, L. Nguyen, T. N. Nguyen, H.-K. Song, and H. Moon, “Lightweight pixel-level semantic segmentation and analysis for sewer defects using deep learning,” Construction and Building Materials, 2023

    work page 2023

  8. [8]

    Co-cracksegment: A new collaborative deep learning framework for pixel-level semantic segmentation of concrete cracks,

    N. F. Alkayem, A. Mayya, L. Shen, X. Zhang, P. G. Asteris, Q. Wang, and M. Cao, “Co-cracksegment: A new collaborative deep learning framework for pixel-level semantic segmentation of concrete cracks,” Mathematics, 2024

    work page 2024

Show all 59 references
  1. [9]

    An automated visual defect segmentation for flat steel surface using deep neural networks,

    D. N. Sabet, M. Zarifi, J. Khoramdel, Y . Borhani, and E. Najafi, “An automated visual defect segmentation for flat steel surface using deep neural networks,” in 2022 12th International Conference on Computer and Knowledge Engineering (ICCKE) , pp. 423–427, 2022

    2022

  2. [10]

    Iter- lunet: deep learning architecture for pixel-wise crack detection in levee systems,

    M. Panta, M. T. Hoque, M. Abdelguerfi, and M. C. Flanagin, “Iter- lunet: deep learning architecture for pixel-wise crack detection in levee systems,” IEEE Access, vol. 11, pp. 12249–12262, 2023

    2023

  3. [11]

    Addressing class imbalance in micro-ct image segmentation: A modified u-net model with pixel-level class weighting,

    S. Mahmoudi, O. Asghari, and J. Boisvert, “Addressing class imbalance in micro-ct image segmentation: A modified u-net model with pixel-level class weighting,” Computers & Geosciences , 2025

    2025

  4. [12]

    Kolmogorov-arnold network autoencoders,

    M. Moradi, S. Panahi, E. Bollt, and Y .-C. Lai, “Kolmogorov-arnold network autoencoders,” ArXiv, 2024

    2024

  5. [13]

    Mof-kan: Kolmogorov- arnold networks for digital discovery of metal-organic frameworks,

    X. Wu, X. Song, Y . Yue, R. Zheng, and J. Jiang, “Mof-kan: Kolmogorov- arnold networks for digital discovery of metal-organic frameworks,” The Journal of Physical Chemistry Letters , vol. 16, pp. 2452–2459, 2025

    2025

  6. [14]

    Kanice: Kolmogorov-arnold networks with interactive convolutional elements,

    M. M. Ferdaus, M. Abdelguerfi, E. Ioup, D. Dobson, K. N. Niles, K. Pathak, and S. Sloan, “Kanice: Kolmogorov-arnold networks with interactive convolutional elements,” in Proceedings of the 4th Interna- tional Conference on AI-ML Systems , pp. 1–10, 2024

    2024

  7. [15]

    Kolmogorov–arnold network for hyperspectral change detection,

    S. T. Seydi, M. Sadegh, and J. Chanussot, “Kolmogorov–arnold network for hyperspectral change detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 63, pp. 1–15, 2025

    2025

  8. [16]

    A white-box deep-learning method for electrical energy system modeling based on kolmogorov-arnold network,

    Z. Zhou, Y . Li, Z. Guo, Z. Yan, and M.-Y . Chow, “A white-box deep-learning method for electrical energy system modeling based on kolmogorov-arnold network,” ArXiv, 2024

    2024

  9. [17]

    Kolmogorov-arnold networks in trans- former attention for low-light image enhancement,

    A. Brateanu and R. Balmez, “Kolmogorov-arnold networks in trans- former attention for low-light image enhancement,” in 2024 Interna- tional Symposium on Electronics and Telecommunications (ISETC) , pp. 1–4, 2024

    2024

  10. [18]

    Stand-alone composite attention network for concrete structural defect classification,

    G. Bhattacharya, N. Puhan, and B. Mandal, “Stand-alone composite attention network for concrete structural defect classification,” IEEE Transactions on Artificial Intelligence , vol. 3, pp. 265–274, 2022

    2022

  11. [19]

    Vibration-based rf-svm for pc structural defect detection and assessment,

    S. Xiao and N. Fomin, “Vibration-based rf-svm for pc structural defect detection and assessment,” E3S Web of Conferences , 2024. JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 12

    2024

  12. [20]

    Localizing structural damage based on auto-regressive with exogenous input model parameters and residuals using a support vector machine based learning approach,

    B. Gunes, “Localizing structural damage based on auto-regressive with exogenous input model parameters and residuals using a support vector machine based learning approach,” Frontiers of Structural and Civil Engineering, 2024

    2024

  13. [21]

    Predictive modeling of structural perfor- mance using machine learning: A comprehensive review,

    A. Rajpoot and S. Tiwari, “Predictive modeling of structural perfor- mance using machine learning: A comprehensive review,” International Research Journal on Advanced Engineering Hub (IRJAEH) , 2025

    2025

  14. [22]

    Machine learning-assisted improved anomaly detection for structural health monitoring,

    S. Samudra, M. Barbosh, and A. Sadhu, “Machine learning-assisted improved anomaly detection for structural health monitoring,” Sensors (Basel, Switzerland), vol. 23, 2023

    2023

  15. [23]

    Assessing the impact of deep learning on grey urban infrastructure systems: A comprehensive review,

    A. Oulahyane, M. Kodad, A. Bouazza, and K. Oulahyane, “Assessing the impact of deep learning on grey urban infrastructure systems: A comprehensive review,” in 2024 International Conference on Decision Aid Sciences and Applications (DASA) , pp. 1–9, 2024

    2024

  16. [24]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 3431–3440, 2015

    2015

  17. [25]

    Comparison of fully convolutional networks and u-net for optic disc and optic cup segmentation,

    Z. Jin, “Comparison of fully convolutional networks and u-net for optic disc and optic cup segmentation,” in ITM Web of Conferences , 2025

    2025

  18. [26]

    Development of semantic segmentation based on deep learn- ing,

    Y . Zhao, “Development of semantic segmentation based on deep learn- ing,” Highlights in Science, Engineering and Technology , 2023

    2023

  19. [27]

    Dual attention u-net with feature infusion: Pushing the boundaries of multiclass defect segmentation,

    R. Alshawi, M. T. Hoque, M. M. Ferdaus, M. Abdelguerfi, K. Niles, K. Prathak, J. Tom, J. Klein, M. Mousa, and J. J. Lopez, “Dual attention u-net with feature infusion: Pushing the boundaries of multiclass defect segmentation,” arXiv preprint arXiv:2312.14053 , 2023

    2023 arXiv

  20. [28]

    Textile defect detection based on multi-proportion spa- tial attention mechanism and channel memory feature fusion network,

    Y . Ji and L. Di, “Textile defect detection based on multi-proportion spa- tial attention mechanism and channel memory feature fusion network,” IET Image Processing , 2023

    2023

  21. [29]

    Pddd-net: Defect detection network based on parallel attention mechanism and dual-channel spatial pyramid pooling,

    T. Sui and J. Wang, “Pddd-net: Defect detection network based on parallel attention mechanism and dual-channel spatial pyramid pooling,” IEEE Access, vol. 11, pp. 141764–141775, 2023

    2023

  22. [30]

    Progressive attention guided recurrent network for salient object detection,

    X. Zhang, T. Wang, J. Qi, H. Lu, and G. Wang, “Progressive attention guided recurrent network for salient object detection,” in Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 714–722, 2018

    2018

  23. [31]

    Wgs yolo dual: A detection model for strip steel surface defects based on attention mechanism and spatial pyramid pooling structure,

    Q. Liu, Y . Zhong, W. Yang, Z. Liu, J. Ruan, and Z. Zeng, “Wgs yolo dual: A detection model for strip steel surface defects based on attention mechanism and spatial pyramid pooling structure,” in 2024 IEEE International Conference on Cognitive Computing and Complex Data (ICCD)...

    2024

  24. [32]

    Adaptive dual attention fusion network for rgb-d surface defect detection,

    X. Jiang, J. Liu, F. Yan, Y . Lu, S. Jin, H. Liu, and M. Xu, “Adaptive dual attention fusion network for rgb-d surface defect detection,” Lecture Notes in Computer Science , pp. 392–406, 2024

    2024

  25. [33]

    Kat to kans: A review of kolmogorov-arnold networks and the neural leap forward,

    D. Basina, J. R. Vishal, A. Choudhary, and B. Chakravarthi, “Kat to kans: A review of kolmogorov-arnold networks and the neural leap forward,” arXiv preprint arXiv:2411.10622 , 2024

    2024 arXiv

  26. [34]

    A sur- vey on kolmogorov-arnold network,

    S. Somvanshi, S. A. Javed, M. M. Islam, D. Pandit, and S. Das, “A sur- vey on kolmogorov-arnold network,” arXiv preprint arXiv:2411.06078 , 2024

    2024 arXiv

  27. [35]

    Kan: Kolmogorov-arnold networks,

    Z. Liu, Y . Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Soljacic, T. Y . Hou, and M. Tegmark, “Kan: Kolmogorov-arnold networks,” ArXiv, vol. abs/2404.19756, 2024

    2024 arXiv

  28. [36]

    Can kan work? exploring the potential of kolmogorov-arnold networks in computer vision,

    Y . Cang, Y . h. Liu, and L. Shi, “Can kan work? exploring the potential of kolmogorov-arnold networks in computer vision,” arXiv preprint, 2024

    2024

  29. [37]

    Medkaformer: When kolmogorov-arnold theorem meets vision transformer for medical image representation,

    G. Wang, Q. Zhu, C. Song, B. Wei, and S. Li, “Medkaformer: When kolmogorov-arnold theorem meets vision transformer for medical image representation,” IEEE Journal of Biomedical and Health Informatics , 2025

    2025

  30. [38]

    Multilevel feature fusion and kan integration for brain tumor segmentation,

    Q. Zeng and Y . Chen, “Multilevel feature fusion and kan integration for brain tumor segmentation,” in 2024 5th International Conference on Machine Learning and Computer Application (ICMLCA) , pp. 211–216, 2024

    2024

  31. [39]

    Kolmogorov-arnold networks for metal surface defect classification,

    M. Krzywda, M. Wermi ´nski, S. Lukasik, and A. H. Gandomi, “Kolmogorov-arnold networks for metal surface defect classification,” arXiv preprint, 2025

    2025

  32. [40]

    Sa- unet: Spatial attention u-net for retinal vessel segmentation,

    C. Guo, M. Szemenyei, Y . Yi, W. Wang, B. Chen, and C. Fan, “Sa- unet: Spatial attention u-net for retinal vessel segmentation,” in 2020 25th international conference on pattern recognition (ICPR) , pp. 1236– 1242, IEEE, 2021

    2020

  33. [41]

    Unet segmentation network of covid-19 ct images with multi-scale attention,

    M. Chen, S. Yi, M. Yang, Z. Yang, and X. Zhang, “Unet segmentation network of covid-19 ct images with multi-scale attention,” Mathematical Biosciences and Engineering , vol. 20, no. 9, pp. 16762–16785, 2023

    2023

  34. [42]

    Gaei-unet: Global attention and elastic interaction u-net for vessel image segmentation,

    R. Xiao and Z. Wan, “Gaei-unet: Global attention and elastic interaction u-net for vessel image segmentation,” arXiv preprint arXiv:2308.08345, 2023

    2023 arXiv

  35. [43]

    U-net: Convolutional net- works for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional net- works for biomedical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention , pp. 234–241, Springer, 2015

    2015

  36. [44]

    Ege-unet: an efficient group enhanced unet for skin lesion segmentation,

    J. Ruan, M. Xie, J. Gao, T. Liu, and Y . Fu, “Ege-unet: an efficient group enhanced unet for skin lesion segmentation,” in International confer- ence on medical image computing and computer-assisted intervention , pp. 481–490, Springer, 2023

    2023

  37. [45]

    U-kan makes strong backbone for medical image segmentation and generation,

    C. Li, X. Liu, W. Li, C. Wang, H. Liu, and Y . Yuan, “U-kan makes strong backbone for medical image segmentation and generation,” arXiv preprint arXiv:2406.02918, 2024

    2024 arXiv

  38. [46]

    Image-based detection of structural defects using hierarchical multi-scale attention,

    C. Benz and V . Rodehorst, “Image-based detection of structural defects using hierarchical multi-scale attention,” in DAGM German Conference on Pattern Recognition, pp. 337–353, Springer, 2022

    2022

  39. [47]

    Dynamic label injection for imbalanced industrial defect segmentation,

    E. Caruso, F. Pelosin, A. Simoni, and M. Boschetti, “Dynamic label injection for imbalanced industrial defect segmentation,” in European Conference on Computer Vision , pp. 50–64, Springer, 2025

    2025

  40. [48]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pp. 2117–2125, 2017

    2017

  41. [49]

    Attention u-net: Learning where to look for the pancreas,

    O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y . Hammerla, B. Kainz, et al. , “Attention u-net: Learning where to look for the pancreas,” in International Conference on Medical Imaging with Deep Learning , pp. 1–10, 2018

    2018

  42. [50]

    Unet++: A nested u-net architecture for medical image segmentation,

    Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, “Unet++: A nested u-net architecture for medical image segmentation,” in Deep learning in medical image analysis and multimodal learning for clinical decision support: 4th international workshop, DLMIA 2018, and 8th ...

    2018

  43. [51]

    Efficientdet: Scalable and efficient object detection,

    M. Tan, R. Pang, and Q. V . Le, “Efficientdet: Scalable and efficient object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 10781–10790, 2020

    2020

  44. [52]

    Unet 3+: A full-scale connected unet for medical im- age segmentation,

    H. Huang, L. Lin, R. Tong, H. Hu, Q. Zhang, Y . Iwamoto, X. Han, Y .-W. Chen, and J. Wu, “Unet 3+: A full-scale connected unet for medical im- age segmentation,” in ICASSP 2020-2020 IEEE international conference on acoustics, speech and signal processing (ICASSP) , pp. 1055–10...

    2020

  45. [53]

    Unext: Mlp-based rapid medical image segmentation network,

    J. M. J. Valanarasu and V . M. Patel, “Unext: Mlp-based rapid medical image segmentation network,” in International conference on medi- cal image computing and computer-assisted intervention , pp. 23–33, Springer, 2022

    2022

  46. [54]

    Rolling-unet: Revi- talizing mlp’s ability to efficiently extract long-distance dependencies for medical image segmentation,

    Y . Liu, H. Zhu, M. Liu, H. Yu, Z. Chen, and J. Gao, “Rolling-unet: Revi- talizing mlp’s ability to efficiently extract long-distance dependencies for medical image segmentation,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, pp. 3819–3827, 2024

    2024

  47. [55]

    H-vit: A hierarchical vision transformer for deformable image registration,

    M. Ghahremani, M. Khateri, B. Jian, B. Wiestler, E. Adeli, and C. Wachinger, “H-vit: A hierarchical vision transformer for deformable image registration,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 11513–11523, 2024

    2024

  48. [56]

    Swin-unet: Unet-like pure transformer for medical image segmenta- tion,

    H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang, “Swin-unet: Unet-like pure transformer for medical image segmenta- tion,” arXiv preprint arXiv:2105.05537 , 2021

    2021 arXiv

  49. [57]

    Mobileunetr: A lightweight end-to-end hybrid vision transformer for efficient medical image segmentation,

    S. Perera, Y . Erzurumlu, D. Gulati, and A. Yilmaz, “Mobileunetr: A lightweight end-to-end hybrid vision transformer for efficient medical image segmentation,” arXiv preprint arXiv:2409.03062 , 2024

    2024 arXiv

  50. [58]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,” Advances in neural information processing systems , vol. 34, pp. 12077–12090, 2021

    2021

  51. [59]

    Fastervit: Fast vision transformers with hierarchical attention,

    A. Hatamizadeh, G. Heinrich, H. Yin, A. Tao, J. M. Alvarez, J. Kautz, and P. Molchanov, “Fastervit: Fast vision transformers with hierarchical attention,” in International Conference on Learning Representations , 2024

    2024

Pith tools

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