Pith. sign in

REVIEW 3 major objections 26 references

A synthetic-only texture network can score near humans on medieval hand discrimination and shows labeled gender is mostly page location.

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 →

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

T0 review

2026-07-13 04:02 UTC pith:53XM3B43

load-bearing objection Useful synthetic-only dense texture tool plus a real human paleographic quiz; gender caution is solid, zero-shot style claim is still preliminary on one corpus. the 3 major comments →

arxiv 2607.09299 v1 pith:53XM3B43 submitted 2026-07-10 cs.CV

TextileNet: Towards Zero-shot Text-style Segmentation of Manuscripts

Anguelos Nicolaou , Antonella Ambrosio , Desiree Di Donato , Georg Vogeler This is my paper
classification cs.CV
keywords writer identificationpaleographytexture segmentationmulti-task learningzero-shot learninghistorical manuscriptssynthetic datagender classification
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.

Paleographers need help on open-set, degraded manuscripts where labeled scribes are scarce and answers are uncertain. TextileNet is a fully convolutional multi-task network trained only on synthetic pseudo-pages so that every pixel gets a dense texture embedding; those embeddings transfer zero-shot to late-medieval Italian account books. The authors also built an 80-question pair-and-triplet visual quiz and ran it anonymously from lay readers to senior paleographers, creating a human baseline for style discrimination on this material. Zero-shot retrieval on the embeddings reaches 67.5 percent overall (72.5 percent on triplets) against a 50 percent chance floor and sits in a similar range to time-constrained humans. On the Naples ASN 1401 register, writer identity is carried by texture while labeled gender is almost entirely explained by page position, so gender-from-handwriting claims on this corpus must be read as scribal-role confounds rather than style differences.

Core claim

Dense pixel-level texture embeddings produced by a multi-task fully convolutional network trained exclusively on synthetic data transfer zero-shot to late-medieval manuscripts and support sub-word style retrieval that approaches human performance on a new paleographic quiz; the same embeddings show that labeled gender in the Naples corpus is largely page location (scribal role), not handwriting style.

What carries the argument

TextileNet: an IUnet backbone that emits 384-dimensional per-pixel embeddings, trained with multi-task 1x1 heads (foreground, font family, font size, character) plus a pixel-level relative-ratio triplet loss that forces hard positive and hard negative geometry over foreground pixels of a synthetic page.

Load-bearing premise

That modern synthetic fonts, English n-grams, and fractal degradations produce embeddings whose nearest-neighbor geometry remains paleographically meaningful on thin, bleed-through, ruled late-medieval Italian pages with no domain adaptation.

What would settle it

Train an identical network on synthetic pages that match the target tradition (period scripts, bleed-through, ruling, stains) and re-run the zero-shot quiz and Naples hand-versus-position ablations; if scores collapse or position no longer dominates gender, the transfer claim fails.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Exploratory paleographic tools can flag contested regions and ambiguous attributions from frozen embeddings without labeled manuscript training.
  • Triplet (two-alternative) framing should replace yes/no pairs as the default for human and automatic style quizzes because it removes the absolute threshold problem.
  • Gender classification results on this corpus should be re-read as scribal-role identification conditioned by institutional layout, not as evidence of intrinsic male/female handwriting differences.
  • A robust unsupervised clustering step on the same embedding field would turn the method into automatic multi-hand page segmentation.

Where Pith is reading between the lines

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

  • If the domain gap can be closed by period-matched synthesis, the same frozen-embedding pipeline could become a default open-set first pass for any poorly labeled archival collection.
  • The spatial-role confound is likely not unique to this monastery; any corpus where gender co-varies with layout or office will need the same position-only ablation before style claims are trusted.
  • Random-projection RGB maps of the embedding field already give historians an immediate visual of style variation without requiring them to trust a black-box classifier.

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

3 major / 0 minor

Summary. The paper introduces TextileNet, a fully convolutional multi-task IUnet trained only on synthetic pseudo-pages (modern fonts, Brown-corpus text, Tormentor degradations) with font-family, font-size, character, and FG/BG heads plus a pixel-level relative-ratio triplet loss (Eq. 1), producing 384-d dense texture embeddings. These embeddings are transferred zero-shot to late-medieval Italian account books (Naples ASN 1401). The authors contribute an 80-item pair/triplet paleographic quiz administered to 49 participants (lay to senior paleographers), establishing a human baseline, and report zero-shot Chamfer retrieval at 67.5% overall (72.5% on triplets). On ASN 1401, frozen-embedding kNN/LR experiments show writer identity is carried by texture (region-level LR ~76%), while labeled gender is largely explained by page location (position-only ≈ TextileNet+position), supporting a caution against naive gender-from-handwriting claims on this corpus.

Significance. If the zero-shot transfer claim holds, the work offers a practical path for open-set, sub-word style analysis in archival paleography without labeled manuscript training data—an important gap given scarce annotations and multi-hand pages. The human quiz is a genuine methodological contribution: a balanced, anonymized baseline for script-style discrimination on late-medieval text that future systems can be scored against. The gender analysis is carefully anti-circular and historically grounded (scribal role vs. handwriting), and the synthetic MTL + triplet-loss design with released code is a reusable building block. Strengths include transparent ablations (Table 2: triplet loss is load-bearing; Table 3: position-only controls), demographic reporting of the quiz, and explicit institutional confounds (Fig. 2). The main significance is therefore both technical (dense zero-shot texture embeddings) and evaluative (human baseline + caution on gender).

major comments (3)
  1. The central zero-shot transfer claim (§3.2–3.4, §4.3–4.4) rests on a single manuscript register (ASN 1401, 8 hands). Synthetic training uses modern font families, English n-grams, and Tormentor fractals; no domain-adaptation ablation or second historical corpus is reported. Without multi-corpus or cross-tradition controls, the 67.5% quiz score and 76.3% region-level hand accuracy cannot be read as evidence of generalizable paleographic style understanding rather than corpus-specific transfer. A second corpus or a controlled domain-gap experiment is needed to support the claim as stated.
  2. §2.3 and Fig. 1 explicitly note that horizontal ruling lines are the most prominent texture differentiation, and bleed-through/cross-outs are major challenges. The zero-shot pipeline (§4.3) averages embeddings over binarized connected components and compares via Chamfer distance, but there is no control that isolates letter-form/ductus style from residual non-style cues (rulings, bleed-through, page-level ink). Because every quiz crop is from a distinct page, page-level intensity confounds are not ruled out. An ablation that masks rulings or evaluates on ruling-free crops would make the style-transfer interpretation load-bearing rather than suggestive.
  3. Table 3 reports high variance on Hand/Region cells (e.g., kNN TextileNet+Pos 44.9±15.2) and only 8 hands; component-level numbers are tighter but still modest (LR TextileNet-only 37.5% component hand accuracy). The paper compares region-level LR hand accuracy to ICFHR 2020 fragment retrieval [25] only loosely. For the claim that embeddings carry genuine writer-specific signal beyond location, a stronger protocol (leave-one-hand-out, open-set retrieval metrics such as mAP, or comparison to a standard handcrafted baseline on the same crops) would better ground the hand-ID result.

Circularity Check

0 steps flagged

No significant circularity: synthetic multi-task training is independent of the real-manuscript quiz and Naples evaluation; gender caution is anti-circular via position-only baselines.

full rationale

TextileNet is trained exclusively on synthetic pseudo-pages (Brown Corpus text, sampled modern fonts, Pango/Cairo rendering, Tormentor fractal degradations) with multi-task cross-entropy plus a pixel-level relative-ratio triplet loss (Eq. 1). Held-out synthetic pixel-error rates (Table 2) and all manuscript results (paleographic quiz §4.2–4.3, zero-shot Chamfer retrieval, Table 3 kNN/LR on frozen embeddings) use data never seen during training and never used to fit any free parameter of the backbone. The gender findings are explicitly anti-circular: position-only classifiers match TextileNet+position accuracy, showing the labeled signal is largely scribal-role location rather than style. Self-citations (Tormentor [19], earlier LBP/writer-ID work by Nicolaou et al.) supply reusable tools or background, not uniqueness theorems or load-bearing premises that force the zero-shot transfer numbers. No claimed prediction reduces by construction to a fitted target; the derivation chain is ordinary empirical transfer learning evaluated on external human and manuscript benchmarks.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central transfer claim rests on standard multi-task and metric-learning practice plus domain assumptions that synthetic modern-font pages with fractal degradations proxy medieval texture, and that connected-component averages plus Chamfer distance fairly probe style. Free parameters are architectural and sampling choices, not fits to manuscript labels. Invented entities are the network, the pixel triplet loss, and the quiz protocol; none claim new physics, only engineered tools with empirical handles on synthetic validation and the quiz.

free parameters (5)
  • embedding_dimension
    IUnet pixel embedding width (192 vs 384) chosen by capacity/memory tradeoff; final model uses 384-d vectors that define the entire zero-shot geometry.
  • max_foreground_pixels_for_triplet
    Hard-positive/negative sampling capped at 50,000 foreground pixels per image; changes the metric loss statistics and training signal.
  • synthetic_style_sampling_distributions
    Font family, size, kerning, alignment, and paragraph-area distributions are hand-designed; they define what “style” the network is forced to separate.
  • triplet_loss_scale_and_epsilon
    Scale factor and ε in Eq. (1) stabilize the relative-ratio loss; not derived from first principles.
  • Chamfer_vs_within_crop_threshold_rule
    Pair questions decide same/different hand by comparing between-crop Chamfer distance to expected within-crop distance from random partitioning—an ad hoc decision rule for the quiz score.
axioms (5)
  • domain assumption Multi-task learning on shared dense features improves generalization to unseen tasks relative to single-task training.
    Invoked in §3.1 citing Caruana; justifies joint FG/font/size/character heads.
  • ad hoc to paper Synthetic document images with modern fonts and Tormentor degradations are a sufficient training domain for zero-shot transfer of texture style to late-medieval manuscripts.
    Core knowledge-transfer premise of §3.1–3.2 and all zero-shot experiments; not independently validated beyond the reported quiz and Naples results.
  • ad hoc to paper Pixels that agree on all synthetic segmentation labels share a style identity for metric learning.
    Defines positives for the pixel-level triplet loss in §3.4.
  • domain assumption Connected-component averages of pixel embeddings plus Chamfer distance are a valid probe of writer style at sub-word granularity.
    Operational definition of zero-shot retrieval in §4.3; authors note it is naive.
  • standard math Standard cross-entropy on multiplexed RGB segmentation maps plus the ratio-style triplet loss yields usable texture embeddings.
    Training objective in §3.3–3.4; ordinary supervised + metric learning mathematics.
invented entities (3)
  • TextileNet (IUnet backbone + 1×1 classification heads) independent evidence
    purpose: Produce dense 384-d texture embeddings and multi-task maps for zero-shot manuscript style analysis.
    New architecture configuration for this problem; independent handle is synthetic pixel-error table and public code intent.
  • Pixel-level Relative Ratio Triplet Loss (Eq. 1) independent evidence
    purpose: Supply a learning signal so font-family/size/character heads do not cancel and the backbone learns style geometry.
    Paper-specific loss form; evidence is the large error drop in Table 2 when the loss is added.
  • Paleographic visual quiz (80 pair/triplet items) independent evidence
    purpose: Establish a human baseline for late-medieval script-style discrimination under anonymity.
    Methodological invention; still open online; scores and misleading items reported.

reviewed 2026-07-13 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TextileNet: Towards Zero-shot Text-style Segmentation of Manuscripts." pith.science (2026). https://pith.science/paper/53XM3B43

@misc{pith2026260709299,
  author       = {Pith},
  title        = {Pith review of: TextileNet: Towards Zero-shot Text-style Segmentation of Manuscripts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/53XM3B43}},
  note         = {Machine review of arXiv:2607.09299}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Automatic writer identification systems have progressed remarkably in recent years, yet their deployment in archival paleography remains limited by the scarcity of labeled training data, open scribe sets, and degraded image quality. We present TextileNet, a fully convolutional multi-task network trained exclusively on synthetic data to produce dense pixel-level texture embeddings, which we transfer zeroshot to historical manuscript analysis. As an original contribution to evaluation methodology, we designed a paleographic visual quiz of 80 pair and triplet questions and administered it to a range from lay participants to senior paleographers under strict anonymity, establishing to our knowledge for the first time a human baseline for script-style discrimination on late medieval text. We employ TextileNet embeddings to perform zero-shot retrieval on sub-word granularity for hand and gender identification. Our experimental results help in building the credibility of TextileNet in the paleographic domain, but more than that demonstrate in experimental terms that the question of gender in handwriting needs to be treated with caution.

Figures

Figures reproduced from arXiv: 2607.09299 by Anguelos Nicolaou, Antonella Ambrosio, Desiree Di Donato, Georg Vogeler.

Figure 1
Figure 1. Figure 1: Indicative annotation of two hands on one page and a qualitative texture map the most prominent texture differentiation. This data highlights how the dif￾Female A Female B Female C Female F Female H All Centered Male D Male E Male G All Centered [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Spatial footprint of every annotated hand in the corpus. Female hands (top row) are the nuns who compiled the register and occupy the main body of each page. Male hands (bottom row) belong to friars or procurators of San Domenico Maggiore who retained a supervisory role and appear as narrow bands at the foot of pages. This struc￾tural asymmetry is an institutional artefact rather than a reflection of handw… view at source ↗
Figure 3
Figure 3. Figure 3: Synthetic pseudo-pages (left) and their multiplexed segmentation maps (right). The segmentation maps encode font family, font size, and character identity in the RGB channels [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: TextileNet Architecture. 3.4 Pixel-level Relative Ratio Triplet Loss Initially, training TextileNet converged to large validation errors on the seg￾mentation heads other than foreground/background; the orthogonality of the font-family and font-size tasks appeared to cause competing gradients to cancel each other out, and the IUnet received no meaningful learning signal from the segmentation heads. We intro… view at source ↗
Figure 5
Figure 5. Figure 5: Indicative TextileNet outputs for all heads on a synthetic validation sample. No-BG maps suppress the background logit before softmax, suggesting an effective receptive field of roughly 40 pixels radius. 4.2 Paleographic User Study As we were unaware of established baselines or even established evaluation pro￾tocols, we designed a user experiment in order to see how humans would perform in this limited vis… view at source ↗
Figure 6
Figure 6. Figure 6: The three easiest questions in the quiz (a), (b), and (c), all answered 100% correctly. The three most misleading questions (d), (e), and (f), having misled 93.78%, 81.63%, and 77.55% of participants respectively. 4.3 Zero-shot and Naive-supervision Writer Identification In order to assess TextileNet’s dense texture embeddings as a generic text-style understanding method, we designed an extremely naive met… view at source ↗
Figure 7
Figure 7. Figure 7: Paleographic quiz question analysis [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Paleographic quiz scoring by demographic; error bars represent 95% confidence intervals estimated via bootstrapping. texture embeddings to respond to the quiz without any training. The method operated in the following steps: 1. The quiz images were fed as wholes to TextileNet. 2. Connected component analysis was performed on each sample crop using the binarization head of TextileNet. 3. For each sample cro… view at source ↗
Figure 9
Figure 9. Figure 9: Zero-shot TextileNet performance on the paleographic quiz: all questions (left), anchor/triplet questions (centre), and yes/no/pair questions (right) [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗

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

26 extracted references · 1 canonical work pages

  1. [1]

    EURASIP Journal on Image and Video Processing2014, 1–10 (2014)

    Al Maadeed, S., Hassaine, A.: Automatic prediction of age, gender, and nationality in offline handwriting. EURASIP Journal on Image and Video Processing2014, 1–10 (2014)

    2014

  2. [2]

    In: Blanton, V., O’Mara, V., Stoop, P

    Ambrosio, A.: Literacy in Neapolitan women’s convents: An example of female handwriting in a late fifteenth-century accounts ledger. In: Blanton, V., O’Mara, V., Stoop, P. (eds.) Nuns’ Literacies in Medieval Europe: The Kansas City Dia- logue. Medieval Women: Texts and Contexts, vol. 27, pp. 89–108. Brepols, Turn- hout (2015).https://doi.org/10.1484/M.MWT...

    work page doi:10.1484/m.mwtc-eb.5.105526 2015

  3. [3]

    In: Ambrosio, A., Vitolo, P

    Ambrosio, A.: Nuns’ writing and account books: A turning point in a late-medieval Neapolitan convent. In: Ambrosio, A., Vitolo, P. (eds.) Medioevo digitale. Docu- menti e archivi. Arte e architettura. Viella, Roma (2024), university of Naples Federico II 16 A. Nicolaou et al

    2024

  4. [4]

    In: Bmvc

    Balntas,V.,Riba,E.,Ponsa,D.,Mikolajczyk,K.:Learninglocalfeaturedescriptors with triplets and shallow convolutional neural networks. In: Bmvc. vol. 1, p. 3 (2016)

    2016

  5. [5]

    Barrow, H.G., Tenenbaum, J.M., Bolles, R.C., Wolf, H.C.: Parametric correspon- dence and chamfer matching: Two new techniques for image matching. Tech. rep., SRI International (1977)

    1977

  6. [6]

    Machine Learning28(1), 41–75 (1997).https: //doi.org/10.1023/A:1007379606734

    Caruana, R.: Multitask learning. Machine Learning28(1), 41–75 (1997).https: //doi.org/10.1023/A:1007379606734

    doi:10.1023/a:1007379606734 1997

  7. [7]

    Pattern Recognition63, 258–267 (2017)

    Christlein, V., Bernecker, D., Hönig, F., Maier, A., Angelopoulou, E.: Writer iden- tification using gmm supervectors and exemplar-svms. Pattern Recognition63, 258–267 (2017)

    2017

  8. [8]

    In: German Conference on Pattern Recognition

    Christlein, V., Bernecker, D., Maier, A., Angelopoulou, E.: Offline writer identifica- tion using convolutional neural network activation features. In: German Conference on Pattern Recognition. pp. 540–552. Springer (2015)

    2015

  9. [9]

    In: 2015 13th International Conference on Document Analysis and Recognition (ICDAR)

    Djeddi, C., Al-Maadeed, S., Gattal, A., Siddiqi, I., Souici-Meslati, L., El Abed, H.: Icdar2015 competition on multi-script writer identification and gender classifica- tion using ‘quwi’database. In: 2015 13th International Conference on Document Analysis and Recognition (ICDAR). pp. 1191–1195. IEEE (2015)

    2015

  10. [10]

    In: 2020 IEEE 30th International Workshop on Machine Learning for Signal Processing (MLSP)

    Etmann, C., Ke, R., Schönlieb, C.B.: iunets: learnable invertible up-and downsam- pling for large-scale inverse problems. In: 2020 IEEE 30th International Workshop on Machine Learning for Signal Processing (MLSP). pp. 1–6. IEEE (2020)

    2020

  11. [11]

    Brown University, Providence, Rhode Island (1979), manual of information to accompany a standard corpus of present-day American English, revised and amplified

    Francis, W.N., Kučera, H.: Brown Corpus Manual. Brown University, Providence, Rhode Island (1979), manual of information to accompany a standard corpus of present-day American English, revised and amplified

    1979

  12. [12]

    Accessed: May 23, 2026

    Google DeepMind: Gemini 1.5 pro (Feb 2024),https://deepmind.google/ technologies/gemini/, multimodal large language model evaluated on zero-shot visual paleographic scribe identification quiz given in pdf form. Accessed: May 23, 2026

    2024

  13. [13]

    In: Proceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision

    Grieggs, S., Henderson, C., Sobecki, S., Gillespie, A., Scheirer, W.: The paleogra- pher’s eye ex machina: Using computer vision to assist humanists in scribal hand identification. In: Proceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision. pp. 7177–7186 (2024)

    2024

  14. [14]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Gupta, A., Vedaldi, A., Zisserman, A.: Synthetic data for text localisation in natu- ral images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2315–2324 (2016)

    2016

  15. [15]

    arXiv preprint arXiv:1406.2227 (2014)

    Jaderberg, M., Simonyan, K., Vedaldi, A., Zisserman, A.: Synthetic data and artificial neural networks for natural scene text recognition. arXiv preprint arXiv:1406.2227 (2014)

    Pith/arXiv arXiv 2014

  16. [16]

    In: AFHA

    Marcelli, A., Parziale, A., Senatore, R., et al.: Some observations on handwriting from a motor learning perspective. In: AFHA. vol. 1022, pp. 6–10. Citeseer (2013)

    2013

  17. [17]

    In: 2016 12th IAPR workshop on document analysis systems (DAS)

    Nicolaou, A., Bagdanov, A.D., Gomez, L., Karatzas, D.: Visual script and language identification. In: 2016 12th IAPR workshop on document analysis systems (DAS). pp. 393–398. IEEE (2016)

    2016

  18. [18]

    In: 2015 13th International Conference on Document Analysis and Recognition (ICDAR)

    Nicolaou, A., Bagdanov, A.D., Liwicki, M., Karatzas, D.: Sparse radial sampling lbp for writer identification. In: 2015 13th International Conference on Document Analysis and Recognition (ICDAR). pp. 716–720. IEEE (2015)

    2015

  19. [19]

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

    Nicolaou, A., Christlein, V., Riba, E., Shi, J., Vogeler, G., Seuret, M.: Tormentor: Deterministic dynamic-path, data augmentations with fractals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2707–2711 (2022) TextileNet: Zero-shot Text-style Segmentation 17

    2022

  20. [20]

    In: 2014 11th IAPR International Workshop on Document Analysis Systems

    Nicolaou, A., Slimane, F., Maergner, V., Liwicki, M.: Local binary patterns for arabic optical font recognition. In: 2014 11th IAPR International Workshop on Document Analysis Systems. pp. 76–80. IEEE (2014)

    2014

  21. [21]

    Sensors22(24), 9650 (2022)

    Rabaev, I., Alkoran, I., Wattad, O., Litvak, M.: Automatic gender and age classi- fication from offline handwriting with bilinear resnet. Sensors22(24), 9650 (2022)

    2022

  22. [22]

    Applied Intelligence53(13), 17154–17177 (2023)

    Rabaev, I., Litvak, M.: Automated gender classification from handwriting: a sys- tematic survey. Applied Intelligence53(13), 17154–17177 (2023)

    2023

  23. [23]

    In: International Conference on Document Analysis and Recognition

    Raven, T., Christlein, V., Fink, G.A.: Interpretable writer recognition via vectors of locally aggregated characters. In: International Conference on Document Analysis and Recognition. pp. 429–445 (2025)

    2025

  24. [24]

    In: international conference on document analysis and recognition

    Raven, T., Matei, A., Fink, G.A.: Self-supervised vision transformers for writer retrieval. In: international conference on document analysis and recognition. pp. 380–396. Springer (2024)

    2024

  25. [25]

    In: 2020 17th International conference on frontiers in handwriting recognition (ICFHR)

    Seuret, M., Nicolaou, A., Stutzmann, D., Maier, A., Christlein, V.: Icfhr 2020 competition on image retrieval for historical handwritten fragments. In: 2020 17th International conference on frontiers in handwriting recognition (ICFHR). pp. 216–

    2020

  26. [26]

    Human Brain Mapping41(10), 2642–2655 (2020)

    Yang, Y., Tam, F., Graham, S.J., Sun, G., Li, J., Gu, C., Tao, R., Wang, N., Bi, H.Y., Zuo, Z.: Men and women differ in the neural basis of handwriting. Human Brain Mapping41(10), 2642–2655 (2020)

    2020

This paper was first reviewed by grok-4.5 on July 13, 2026.