pith. sign in

arxiv: 2606.22625 · v1 · pith:PF77LSQFnew · submitted 2026-06-21 · 💻 cs.CV

DR-Mamba: Automatic Inference-Time Domain Adaptation for Document Image Binarization via Sample-Conditioned Detail-Background Suppression

Sheng-Wei Chan , Jen-Shiun Chiang This is my paper

Pith reviewed 2026-06-26 10:53 UTC · model grok-4.3

classification 💻 cs.CV
keywords document image binarizationinference-time adaptationdomain shiftMambasubtractive suppressiondegraded documentscross-domain robustness
0
0 comments X

The pith

Input-dependent subtractive gates adapt document binarization to each sample's degradation in one forward pass without labels or updates.

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

The paper establishes that automatic inference-time domain adaptation for degraded document binarization is possible by reinterpreting Mamba scanning as parallel fast detail and slow background routes whose outputs are combined by an input-dependent subtractive gate. This gate explicitly suppresses spatially persistent degradation responses rather than fusing features by addition or concatenation, all within a single forward pass that requires no target-domain labels, no fine-tuning, and no parameter updates. Full-resolution detail-guided reconstruction and thin-stroke-aware supervision are added to recover fine strokes lost in downsampling. A sympathetic reader would care because standard learning-based binarizers become unstable under shifts such as paper aging, bleed-through, or uneven illumination, and the proposed mechanism claims to stabilize separation on unseen domains.

Core claim

DR-Mamba performs automatic inference-time domain adaptation for document image binarization by using input-dependent gates to suppress background interference through a fast-slow route modeling of Mamba scanning, achieving improved cross-domain robustness without requiring target labels, fine-tuning, or test-time updates.

What carries the argument

The input-dependent subtractive gate that integrates a fast detail route for local stroke structures with a slow background route for persistent degradation responses.

If this is right

  • Per-document subtractive suppression yields stronger cross-domain results than additive or concatenative fusion.
  • Particularly large gains appear on the most severely degraded held-out folds under leave-one-year-out evaluation.
  • Full-resolution detail-guided reconstruction recovers thin strokes that would otherwise be lost to downsampling.
  • The entire adaptation occurs inside one forward pass with no gradient steps or auxiliary data required at inference.

Where Pith is reading between the lines

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

  • The subtractive-gate design could be tested on other degradation-sensitive vision tasks such as low-light enhancement or shadow removal to check whether the fast-slow separation generalizes beyond binarization.
  • Replacing the Mamba backbone with a different selective state-space model would reveal whether the reported robustness stems primarily from the gate or from the specific scanning mechanism.
  • Applying the same per-location suppression at multiple scales might further improve performance on documents containing both fine text and large stained regions.
  • The leave-one-year-out protocol on DIBCO-style data leaves open whether similar gains would hold on real archival collections whose degradation distributions differ from the benchmark years.

Load-bearing premise

The input-dependent subtractive gate can reliably separate detail from background degradation using only the single input sample and no target labels or parameter updates.

What would settle it

If a held-out degradation domain shows no accuracy gain for DR-Mamba over a non-adaptive baseline that uses the same Mamba backbone without the subtractive gate, the central claim of improved robustness via per-document per-location suppression would be falsified.

Figures

Figures reproduced from arXiv: 2606.22625 by Jen-Shiun Chiang, Sheng-Wei Chan.

Figure 1
Figure 1. Figure 1: Simplified overview of the proposed DR-Mamba framework. A ConvNeXt-Tiny backbone extracts hierarchical document features, which are restored by a multi-scale decoder. The proposed DR-Mamba block models the decoded representation through a fast detail route and a slow background route, followed by adaptive subtractive sup￾pression to reduce degradation interference. Full-resolution edge cues are further int… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison on challenging held-out documents. DR-Mamba better suppresses background degradation while preserving thin strokes and stroke continuity [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
read the original abstract

Degraded document image binarization is sensitive to domain shifts caused by paper aging, bleed-through, stains, shadows, and uneven illumination, and the foreground-background separation of recent learning-based methods can become unstable on unseen degradation domains. We propose DR-Mamba, a sample-conditioned detail-background suppression framework that performs automatic inference-time domain adaptation for document image binarization. Unlike test-time adaptation methods that require gradient updates or auxiliary data at inference, DR-Mamba adapts to each input document through input-dependent gates within a single forward pass, requiring no target-domain labels, no fine-tuning, and no test-time parameter updates. Instead of using Mamba-style selective scanning as a single generic feature path, DR-Mamba reinterprets it as fast-slow route modeling: a fast detail route captures local stroke structures, while a slow background route accumulates spatially persistent degradation responses. The two routes are integrated through an input-dependent subtractive gate that explicitly suppresses background interference rather than fusing features by addition or concatenation. We further add full-resolution detail-guided reconstruction and thin-stroke-aware supervision to recover fine strokes lost during downsampling. Evaluated under a leave-one-year-out protocol on DIBCO-style benchmarks, where each held-out year is treated as an unseen degradation domain, DR-Mamba shows that per-document, per-location subtractive suppression improves cross-domain robustness, with particularly strong performance on the most severely degraded held-out fold.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes DR-Mamba, a sample-conditioned framework for document image binarization that reinterprets Mamba selective scanning as parallel fast-detail and slow-background routes whose outputs are combined by an input-dependent subtractive gate. This enables automatic inference-time domain adaptation in a single forward pass with no target labels, no fine-tuning, and no parameter updates. Additional components include full-resolution detail-guided reconstruction and thin-stroke-aware supervision. Under a leave-one-year-out protocol on DIBCO-style benchmarks treating each held-out year as an unseen degradation domain, the method claims improved cross-domain robustness, with strongest gains on the most severely degraded folds.

Significance. If the input-dependent subtractive gate reliably isolates local stroke detail from spatially persistent degradation using only the single input sample, the approach would offer a lightweight, label-free alternative to existing test-time adaptation techniques that require gradient steps or auxiliary data. The emphasis on per-document, per-location suppression without retraining addresses a practical need in historical document processing where degradation varies across collections. The absence of any auxiliary diagnostic for the claimed routing, however, leaves the central adaptation mechanism unverified.

major comments (2)
  1. [Method description of subtractive gate and route modeling] The central mechanism reinterprets Mamba scanning as fast-slow routes combined by an input-dependent subtractive gate, yet the manuscript supplies no auxiliary loss, route-specific regularization, or diagnostic visualization to confirm that the gate performs the claimed per-location background suppression rather than some other internal computation. Training uses only the final binarization loss plus thin-stroke supervision; any failure of this implicit separation would directly undermine the cross-domain robustness claim on held-out year folds.
  2. [Abstract and Evaluation] The abstract asserts performance gains under the leave-one-year-out protocol, but the supplied text contains no quantitative metrics, ablation tables, or implementation details that would allow assessment of the cross-domain claims or the contribution of the subtractive gate versus the reconstruction and supervision terms.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on DR-Mamba. We address each major comment below and outline revisions to strengthen the verification of the proposed mechanism and the presentation of results.

read point-by-point responses

  1. Referee: [Method description of subtractive gate and route modeling] The central mechanism reinterprets Mamba scanning as fast-slow routes combined by an input-dependent subtractive gate, yet the manuscript supplies no auxiliary loss, route-specific regularization, or diagnostic visualization to confirm that the gate performs the claimed per-location background suppression rather than some other internal computation. Training uses only the final binarization loss plus thin-stroke supervision; any failure of this implicit separation would directly undermine the cross-domain robustness claim on held-out year folds.

    Authors: We acknowledge that the current training objective relies solely on the final binarization loss and thin-stroke supervision without explicit route-specific terms or diagnostics. This leaves the internal behavior of the subtractive gate as an implicit outcome. To address this, we will add diagnostic visualizations of the gate outputs and ablation experiments that isolate the gate's contribution in the revised manuscript. revision: yes

  2. Referee: [Abstract and Evaluation] The abstract asserts performance gains under the leave-one-year-out protocol, but the supplied text contains no quantitative metrics, ablation tables, or implementation details that would allow assessment of the cross-domain claims or the contribution of the subtractive gate versus the reconstruction and supervision terms.

    Authors: The full manuscript includes quantitative tables, ablations, and implementation details in the Experiments section. However, the abstract itself does not report specific numbers. We will revise the abstract to include key metrics from the leave-one-year-out results and ensure clearer linkage between components and performance gains. revision: partial

Circularity Check

0 steps flagged

No circularity; architectural design is independently specified and empirically evaluated

full rationale

The paper presents a neural architecture (DR-Mamba) that reinterprets Mamba selective scanning as fast-slow routes combined by an input-dependent subtractive gate, trained end-to-end with standard binarization loss plus thin-stroke supervision. No equations, derivations, or first-principles results are shown that reduce any claimed prediction or output to the inputs by construction. No self-citations are invoked to justify uniqueness theorems, import ansatzes, or load-bear the central mechanism. The cross-domain robustness claim rests on leave-one-year-out empirical results rather than tautological definitions or fitted-input renaming. This is the common case of a self-contained empirical method paper with no detectable circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only abstract available; no equations, training details, or model components visible to identify free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5792 in / 1027 out tokens · 31709 ms · 2026-06-26T10:53:16.108286+00:00 · methodology

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

32 extracted references · 17 canonical work pages

  1. [1]

    arXiv preprint arXiv:2312.03568 (2023)

    Biswas, R., Roy, S.K., Wang, N., Pal, U., Huang, G.B.: Docbinformer: A two- level transformer network for effective document image binarization. arXiv preprint arXiv:2312.03568 (2023)

    arXiv 2023

  2. [2]

    In: Pro- ceedings of the Fourteenth Indian Conference on Computer Vision, Graphics and Image Processing

    Biswas, R., Sarkhel, S., Roy, S.K., Pal, U.: Transdocunet: A transformer- based unet architecture for degraded document image binarization. In: Pro- ceedings of the Fourteenth Indian Conference on Computer Vision, Graphics and Image Processing. pp. 1–9. Association for Computing Machinery (2023). https://doi.org/10.1145/3627631.3627639

    work page doi:10.1145/3627631.3627639 2023

  3. [3]

    arXiv preprint arXiv:2404.05669 (2024)

    Cicchetti, G., Comminiello, D.: Naf-dpm: A nonlinear activation-free diffusion probabilistic model for document enhancement. arXiv preprint arXiv:2404.05669 (2024)

    arXiv 2024

  4. [4]

    IEEE Signal Processing Letters27, 1090–1094 (2020)

    De, R., Chakraborty, A., Sarkar, R.: Document image binarization using dual dis- criminator generative adversarial networks. IEEE Signal Processing Letters27, 1090–1094 (2020). https://doi.org/10.1109/LSP.2020.3003828

    work page doi:10.1109/lsp.2020.3003828 2020

  5. [5]

    , author Dong, W

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition. pp. 248–255 (2009). https://doi.org/10.1109/CVPR.2009.5206848

    work page doi:10.1109/cvpr.2009.5206848 2009

  6. [6]

    Gatos, K

    Gatos, B., Ntirogiannis, K., Pratikakis, I.: Icdar 2009 document image bi- narization contest (dibco 2009). In: Proceedings of the International Con- ference on Document Analysis and Recognition. pp. 1375–1382 (2009). https://doi.org/10.1109/ICDAR.2009.246

    work page doi:10.1109/icdar.2009.246 2009

  7. [7]

    Pattern Recognition39(3), 317–327 (2006)

    Gatos, B., Pratikakis, I., Perantonis, S.J.: Adaptive degraded docu- ment image binarization. Pattern Recognition39(3), 317–327 (2006). https://doi.org/10.1016/j.patcog.2005.09.010

    work page doi:10.1016/j.patcog.2005.09.010 2006

  8. [8]

    arXiv preprint arXiv:2312.00752 (2023) 16 S.-W

    Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023) 16 S.-W. Chan et al

    Pith/arXiv arXiv 2023

  9. [9]

    In: International Conference on Learning Representations (2022)

    Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. In: International Conference on Learning Representations (2022)

    2022

  10. [10]

    Pattern Recognition91, 379–390 (2019)

    He, S., Schomaker, L.: Deepotsu: Document enhancement and binariza- tion using iterative deep learning. Pattern Recognition91, 379–390 (2019). https://doi.org/10.1016/j.patcog.2019.01.025

    work page doi:10.1016/j.patcog.2019.01.025 2019

  11. [11]

    arXiv preprint arXiv:2512.14114 (2025)

    Ju, R.Y., Wong, K., Jin, Y., Chiang, J.S.: Mfe-gan: Efficient gan-based frame- work for document image enhancement and binarization with multi-scale feature extraction. arXiv preprint arXiv:2512.14114 (2025)

    arXiv 2025

  12. [12]

    In: International Conference on Medical Imaging with Deep Learning

    Kervadec, H., Bouchtiba, J., Desrosiers, C., Granger, E., Dolz, J., Ayed, I.B.: Boundary loss for highly unbalanced segmentation. In: International Conference on Medical Imaging with Deep Learning. pp. 285–296 (2019)

    2019

  13. [13]

    In: Advances in Neural Information Processing Systems

    Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Jiao, J., Liu, Y.: Vmamba: Visual state space model. In: Advances in Neural Information Processing Systems. vol. 37 (2024)

    2024

  14. [14]

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

    Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11976–11986 (2022)

    2022

  15. [15]

    In: International Conference on Learning Representations (2019)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)

    2019

  16. [16]

    In: International Conference on Learning Representations (2018)

    Micikevicius, P., Narang, S., Alben, J., Diamos, G., Elsen, E., Garcia, D., Ginsburg, B., Houston, M., Kuchaiev, O., Venkatesh, G., Wu, H.: Mixed precision training. In: International Conference on Learning Representations (2018)

    2018

  17. [17]

    IEEE Transactions on Systems, Man, and Cybernetics , author =

    Otsu, N.: A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics9(1), 62–66 (1979). https://doi.org/10.1109/TSMC.1979.4310076

    work page doi:10.1109/tsmc.1979.4310076 1979

  18. [18]

    In: Advances in Neural Information Processing Systems

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Informat...

    2019

  19. [19]

    In: Proceedings of the International Conference on Frontiers in Handwriting Recognition

    Pratikakis, I., Gatos, B., Ntirogiannis, K.: H-dibco 2010: Handwritten doc- ument image binarization competition. In: Proceedings of the International Conference on Frontiers in Handwriting Recognition. pp. 727–732 (2010). https://doi.org/10.1109/ICFHR.2010.118

    work page doi:10.1109/icfhr.2010.118 2010

  20. [20]

    Pratikakis, B

    Pratikakis, I., Zagoris, K., Barlas, G., Gatos, B.: Icdar 2013 document im- age binarization contest (dibco 2013). In: Proceedings of the International Conference on Document Analysis and Recognition. pp. 1471–1476 (2013). https://doi.org/10.1109/ICDAR.2013.219

    work page doi:10.1109/icdar.2013.219 2013

  21. [21]

    1395–1403

    Pratikakis, I., Zagoris, K., Barlas, G., Gatos, B.: Icdar 2017 competition on document image binarization (dibco 2017). In: Proceedings of the Interna- tional Conference on Document Analysis and Recognition. pp. 1395–1403 (2017). https://doi.org/10.1109/ICDAR.2017.228

    work page doi:10.1109/icdar.2017.228 2017

  22. [22]

    In: 2018 16th International Conference on Frontiers in Handwriting Recognition

    Pratikakis, I., Zagoris, K., Kaddas, P., Gatos, B.: Icfhr 2018 competition on hand- written document image binarization (h-dibco 2018). In: Proceedings of the Inter- national Conference on Frontiers in Handwriting Recognition. pp. 489–493 (2018). https://doi.org/10.1109/ICFHR-2018.2018.00091

    work page doi:10.1109/icfhr-2018.2018.00091 2018

  23. [23]

    In: International Workshop on Machine Learning in Medical Imaging

    Salehi, S.S.M., Erdogmus, D., Gholipour, A.: Tversky loss function for image seg- mentation using 3d fully convolutional deep networks. In: International Workshop on Machine Learning in Medical Imaging. pp. 379–387. Springer (2017) DR-Mamba for Document Image Binarization 17

    2017

  24. [24]

    Pattern Recognition33(2), 225–236 (2000)

    Sauvola, J., Pietikainen, M.: Adaptive document image binarization. Pattern Recognition33(2), 225–236 (2000). https://doi.org/10.1016/S0031-3203(99)00055- 2

    work page doi:10.1016/s0031-3203(99)00055- 2000

  25. [25]

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

    Shit, S., Paetzold, J.C., Sekuboyina, A., Ezhov, I., Unger, A., Zhylka, A., Pluim, J.P.W., Bauer, U., Menze, B.H.: cldice: A novel topology-preserving loss function for tubular structure segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16560–16569 (2021)

    2021

  26. [26]

    arXiv preprint arXiv:2201.10252 (2022)

    Souibgui, M.A., Biswas, S., Jemni, S.K., Kessentini, Y., Fornés, A., Lladós, J., Pal, U.: Docentr: An end-to-end document image enhancement transformer. arXiv preprint arXiv:2201.10252 (2022)

    arXiv 2022

  27. [27]

    IEEE Transactions on Image Processing22(4), 1408– 1417 (2013)

    Su, B., Lu, S., Tan, C.L.: A robust document image binarization technique for degraded document images. IEEE Transactions on Image Processing22(4), 1408– 1417 (2013). https://doi.org/10.1109/TIP.2012.2231089

    work page doi:10.1109/tip.2012.2231089 2013

  28. [28]

    In: Proceedings of the International Con- ference on Document Analysis and Recognition

    Tensmeyer, C., Martinez, T.: Document image binarization with fully convolutional neural networks. In: Proceedings of the International Con- ference on Document Analysis and Recognition. pp. 99–104 (2017). https://doi.org/10.1109/ICDAR.2017.25

    work page doi:10.1109/icdar.2017.25 2017

  29. [29]

    Information Fusion93, 159–173 (2023)

    Yang, M., Xu, S.: A novel degraded document binarization model through vision transformer network. Information Fusion93, 159–173 (2023). https://doi.org/10.1016/j.inffus.2023.01.007

    work page doi:10.1016/j.inffus.2023.01.007 2023

  30. [30]

    In: Proceedings of the 31st ACM International Conference on Multimedia

    Yang, Z., Liu, B., Xiong, Y., Yi, L., Wu, G., Tang, X., Liu, Z., Zhou, J., Zhang, X.: Docdiff: Document enhancement via residual diffusion models. In: Proceedings of the 31st ACM International Conference on Multimedia. Association for Computing Machinery (2023). https://doi.org/10.1145/3581783.3611730

    work page doi:10.1145/3581783.3611730 2023

  31. [31]

    Pattern Recog- nition96, 106968 (2019)

    Zhao, J., Shi, C., Jia, F., Wang, Y., Xiao, B.: Document image binarization with cascaded generators of conditional generative adversarial networks. Pattern Recog- nition96, 106968 (2019). https://doi.org/10.1016/j.patcog.2019.106968

    work page doi:10.1016/j.patcog.2019.106968 2019

  32. [32]

    In: Proceedings of the International Conference on Machine Learning

    Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. In: Proceedings of the International Conference on Machine Learning. pp. 62429–62442 (2024)

    2024