Out-of-Length Scene Text Recognition: A Two-Axis Diagnosis and a Training-Free Fix

John Zelek; Zobeir Raisi

REVIEW 3 major objections 4 minor 47 references

Out-of-Length Scene Text Recognition: A Two-Axis Diagnosis and a Training-Free Fix

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read This paper argues that out-of-length scene text recognition is two coupled extrapolation failures, with encoder width dominant, and that a training-free crop-decode-stitch procedure lets unmodified word-level checkpoints match and partly be

desk verdict Chunk-decode-stitch is a genuinely useful, training-free fix for long-text STR, and the two-axis diagnosis is honest and well-evidenced; the main soft spots are the published-baseline comparison and the unquantified over-merge failure mode. read the letter →

arxiv 2607.23194 v1 pith:M3MW27MB submitted 2026-07-25 cs.CV

Zobeir Raisi , John Zelek This is my paper · ORCID

classification cs.CV

keywords scenetextrecognitionlengthextrapolationencoderwidthaxisdecodertimerotarypositionembeddingrescalingchunkeddecodingedit-distancestitchinglongbenchmark

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

Most scene text recognizers are trained on word crops of at most 25 characters, so a longer image forces two things to extrapolate at once: the encoder sees more width tokens, and the decoder must emit more steps. The paper separates these axes and argues the encoder's width axis is the dominant failure, behaving like a representation leaving its training distribution rather than a positional code running out of range; resolution-preserving rotary rescaling does nothing, while interpolation-style rescalings recover only a few CER points. The residual gap is decoding-side, and the paper closes it with zero training by slicing the long image into overlapping crops at the model's training width, decoding each independently and in-distribution, and stitching the reads with geometry-anchored edit-distance alignment. On the Long Text Benchmark this reaches 42.79–43.05% bucket-average word accuracy, matching the published state of the art and beating it by 11–12 points on the hardest bucket, at wall-clock parity with plain decoding. A sympathetic reader would take this as evidence that the long-text problem can be moved largely out of the network, into image-space cropping and string alignment.

What carries the argument

Chunk-decode-stitch, an inference-time algorithm with no learned parameters. The long image is aspect-preserving resized to the model's training height, sliced into overlapping crops at exactly the training width (default stride half the width), each crop upsampled if narrower than training width, and all crops decoded in one batched forward pass by the frozen greedy decoder with a short per-crop budget, so every decode is in-distribution on both extrapolation axes. The per-crop reads are then merged left-to-right by geometry-anchored edit-distance alignment: for each adjacent pair the expected pixel overlap is translated into a candidate character overlap, a small window of overlap lengths

What would settle it

Count how often the stitcher over-merges at repeated substrings on the hardest Long Text Benchmark bucket by running chunk-decode-stitch on images whose true overlap boundaries are known from the ground truth; if spurious merges occur in more than a small percentage of the 56+ bucket, the claimed 11–12 point margin over the published baseline on that bucket would shrink correspondingly, even though the two-axis diagnosis and per-crop reads would remain intact.

Watch

Extended reading notes

Core claim

The paper's central claim has a diagnostic and a practical half. Diagnostically, out-of-length failure is not one problem but two coupled ones—the encoder's width axis and the decoder's time axis—and they fail unequally: width-only extrapolation already does most of the damage, even when only the first 25 characters need to be read. Representation-side fixes confirm the asymmetry: a resolution-preserving rotary rescaling is a clean null, and interpolation-style rescalings recover only a small positional-geometry component, while fine-tuning recovers more CER but leaves exact-match word accuracy near zero. The remaining gap is shown to be decoding-side, since neither rescaling the decoder's p

Load-bearing premise

The headline accuracy rests on the assumption that each short crop's read is faithful enough within its window, and the geometry-anchored edit-distance matcher reliable enough, that the stitcher recovers the true overlap and does not regularly join the wrong pieces when a short string repeats; the paper documents this over-merge failure only qualitatively.

Editorial extensions

If this is right

Any existing word-level checkpoint—including public ones that score zero on long text—can be turned into a long-text reader with no retraining, new data, or architecture change, at wall-clock parity with plain decoding.
The dominance of the width axis redirects future length-robustness work toward encoders that stay well-conditioned as token count grows, away from positional-encoding tweaks.
The decoder-axis gap is not a positional-encoding problem, so purpose-built decoders that re-decode or stitch remain the right target when a training budget exists, and any new long-text method should beat this training-free baseline.
Fine-tuning for width-axis robustness and chunked decoding are near-orthogonal: once chunking feeds only in-distribution crops, width-axis fine-tuning has nothing left to correct and mildly hurts.
The method is cost-neutral because crops are batched into one parallel forward pass per decode step, while plain decoding advances autoregressively one step at a time.

Reading between the lines

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

If the width-axis diagnosis generalizes to other 2D visual encoders, length extrapolation in document understanding, full-page OCR, and dense captioning may be best addressed by giving encoders more token-count variety at training time, not by positional-frequency rescaling.
The known but unquantified over-merge failure at repeated substrings suggests a confidence-weighted or lightly learned stitcher, or a two-dimensional crop grid, would extend the method to multi-line and vertical layouts that the current horizontal window does not handle.
Because chunking makes width-axis fine-tuning unnecessary, one could test whether a training-time curriculum that always presents short, in-distribution crops produces a checkpoint that is long-text-capable with plain decoding, removing the stitching step entirely.
The contrast between the null result of resolution-preserving rotary rescaling here and its reported success in language modelling is a transferable hypothesis: measure on other 2D anisotropic encoders whether interpolation-style rescaling, rather than resolution-preserving rescaling, is generally the only effective positional correction.

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.

Desk Editor's Note

read the letter

Colleague,

Read the Raisi-Zelek paper. Short version: it's worth your time. The headline result — that you can take an off-the-shelf word-level STR checkpoint, slice long images into overlapping 128px crops, decode each greedily, and stitch the reads with edit-distance alignment, matching or beating the purpose-built SMTR on LTB without training — is real as far as I can tell. They also make a clean point that encoder width, not decoder length, is the main culprit, and they back it with a careful axis-by-axis decomposition.

What the paper does well: the evaluation is unusually honest. They report the NTK rescaling as a null result, they show the narrow-long synthetic stream made things worse, they sweep stride, phase, and base-training seed, and they release code. The PARSeq transfer row is a nice touch, even though it mixes corpora — they flag that clearly. The synthetic curved-text test is a good addition.

Where I'd push back: the headline comparison against SMTR, LISTER etc. relies on published numbers rather than a same-harness re-run. That's conventional in this field, but it leaves room for hidden differences in decoding protocol or scoring. The 30-character per-crop budget is also slightly above the stated 25-char training cap; minor, but the 'in-distribution' claim should be softened. The bigger substantive concern is the stitch stage: the paper admits in §5 that over-merging at repeated substrings is not quantified. Since the L≥56 margin (36.84 vs 25.56) is exactly where stitching errors would bite, a few percent of over-merges would shrink that margin materially. I'd want that measured as an error category before trusting the 11-point claim. Also, the width-axis dominance result is partly confounded by task difficulty — the width-only condition only needs the first 25 characters, so higher CER there isn't quite proof that width matters more. The authors acknowledge this, so it's a nuance not a fatal flaw.

Bottom line: this is a solid empirical paper with a useful, deployable result. The central claim is probably correct, but the baseline comparison and the over-merge quantification need tightening. It deserves a serious referee, and I'd engage with it in my own work.

Best.

Referee Report

3 major / 4 minor

Summary. The paper studies why word-level STR models (trained on <=25-character crops) fail on longer text. It decomposes out-of-length failure into an encoder-width axis and a decoder-time axis, and reports that encoder-width extrapolation dominates (Table 1). Representation-side fixes -- NTK/PI/YaRN rescaling and weighted fine-tuning -- give partial CER gains, but word accuracy remains near zero, localizing the residual gap to decoding. The paper then proposes chunk-decode-stitch: overlapping crops at training width are decoded independently and stitched by geometry-anchored edit-distance alignment. On LTB, the method reaches 42.79-43.05% bucket-average word accuracy on two base checkpoints, comparable to or above published SMTR (41.57), with a larger margin on the hardest bucket; applied to the public PARSeq checkpoint it reaches 47.11%, at wall-clock parity with plain decoding. The authors include robustness sweeps and several honest negative results.

Significance. If the headline numbers hold, this is an important result: a training-free, checkpoint-agnostic inference procedure can turn an unmodified word-level STR model into a long-text reader, and the two-axis diagnosis redirects future work from positional-frequency tweaks toward representation robustness. The paper is unusually careful in several respects: a controlled synthetic length curve, stride/phase/base-seed sweeps, reproducible code release, and honest reporting of negative results (NTK null, the failed third fine-tuning stream, the fine-tuning-seed spread for chunking). The main risks are the unquantified stitch-stage over-merge failure and the reliance on published baseline numbers rather than a same-harness re-run; both are testable and should be resolved before the central empirical claim is accepted.

major comments (3)

[§3.5.2, §5, Table 7] The stitcher's overlap selection uses normalized edit-distance similarity over candidate overlap lengths k, with fallback below a minimum-similarity threshold. When a repeated substring occurs at a crop boundary, a wrong k can outscore the true overlap, causing non-adjacent text to be merged. Since LTB is scored by exact word match, every such merge becomes a word-level miss. Section 5 acknowledges this over-merging failure but does not quantify it, and the only manual inspection (12 hardest-bucket examples, §4.10) cannot bound its rate. The L≥56 margin (36.84 vs. SMTR 25.56) is a headline contribution; if over-merges occur on even a few percent of that bucket the margin shrinks materially. Please add an error-category breakdown on LTB separating stitcher over-merges from crop-read errors, and a targeted repeated-substring synthetic stress test, with the minimum-similarity threshold swep
[§4.10, Table 7] The claim of matching or beating the 'published state of the art' rests entirely on published baseline numbers from Du et al. [10] rather than a same-harness re-run; resizing, decode protocol, and evaluation details may differ across runs. The public-PARSeq row (47.11) is additionally trained on a different corpus, a point the text acknowledges but which makes the 5.5-point advantage over SMTR a checkpoint-agnosticism demonstration rather than a controlled comparison. Please either re-run SMTR, LISTER, FocalSVTR, and AR-STR under the same harness and report them in Table 7, or explicitly frame the contribution as 'matching the best published number' and add uncertainty estimates. The latter is the minimum needed to avoid implying a controlled comparison.
[§1, §5, §S3] The paper motivates the problem with 'storefront signage, multi-word phrases, dense captions' and says the method 'closes' the out-of-length failure, but the evaluation is confined to LTB's mostly regular horizontal text plus synthetic long-curved text. Dense multi-line and vertical layouts are untested, and the horizontal sliding-window prior is likely to fail or require a two-dimensional crop grid in exactly those settings. The authors list this as a limitation in §5, but it cuts against the abstract's broad claim. Please add a small dense/vertical subset, even synthetic, to bound performance, or explicitly restrict the claims to regular horizontal long text.

minor comments (4)

[Fig. 1, §3.5.2, contribution 4] The paper repeatedly calls crop stride the 'only hyperparameter', but §3.5.2 also has a minimum-similarity threshold for the overlap match, and there is a per-crop decode budget and a width-gating threshold. Please reconcile this statement and report a threshold sweep, even if the result is insensitive.
[§3.5.3, Table 7] The width-gating rule is stated in the method but is not applied in the main LTB table. State whether any LTB images fall below the width threshold and whether applying the gate changes the reported numbers.
[Supplementary S1] The CER values exceeding 100% ('repetition-loop signature') should be explained in the text, since corpus-level CER is normally bounded by 100 only when deletions are unrestricted; insertions can push it above. A one-sentence clarification or a capped display would help.
[Supplementary S2] The fine-tuning-seed sweep uses only three seeds and shows a 9.3-point A-Avg spread. The claim that 'no seed helps' is sensitive to this small n; consider adding more seeds or reporting a confidence interval.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the LTB accuracy of chunk-decode-stitch is an external empirical measurement, and the two-axis diagnosis has independent experimental content.

full rationale

I walked the derivation chain and found no step in which a claimed prediction or diagnosis is equivalent to its input by construction. The central headline number (42.79–43.05% bucket-average word accuracy on LTB) is measured on an external real-image benchmark against published baselines; it is not derived from the two-axis diagnosis or from the parameters of the method. The chunk-decode-stitch stitcher picks overlap by normalized edit-distance similarity between per-crop reads with a geometric fallback threshold; this is an algorithmic alignment procedure, not a parameter fitted to the LTB labels, and its output is not the same object as any fitted input. The two-axis decomposition is defined by ablations that hold one axis at its training condition (Sections 3.1, 4.3), and the claim that encoder width dominates decoder length is a measured comparison on LTB and a synthetic curve, not a definitional consequence. The conclusion that the residual gap is decoding-side rests on multiple controlled interventions (decoder-table rescaling, a rotary decoder, fine-tuning recipes) and is an interpretive inference from those experiments rather than a reduction. The base architecture is cited from the first author's prior work [14], but the paper specifies all architecture details in Sections 3.1 and 4.2, reports its standard-benchmark accuracy empirically, and the results are reproduced across two independent base-training seeds; no uniqueness theorem or unverified ansatz is imported from that self-citation. Related-work self-citations [41–43] are not load-bearing. The paper also explicitly acknowledges real limitations: the stitcher's over-merging failure at repeated substrings is 'document[ed] qualitatively but do[es] not quantify as its own error category' (Section 5), and dense multi-line or vertical layouts are untested. These are correctness and robustness caveats about whether the headline accuracy would survive harder layouts or error reclassification; they do not make the derivation circular, because the benchmark result remains a falsifiable external measurement. No equation was found that equals another by construction, and no fitted input was renamed as a prediction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central results depend on benchmark conventions, hyperparameters, and standard math, not on new postulated entities. The paper introduces no new particle, force, dimension, or conserved quantity; its invention is algorithmic (chunk-decode-stitch), which is fully specified as a procedure rather than an entity.

free parameters (6)

crop stride s = 64 px (50% overlap)
The single hyperparameter of chunk-decode-stitch; chosen from a 32–112 px sweep showing a plateau at ≥50% overlap (Fig. 4). Not fit per-example.
crop-grid phase p = 0 (default; swept 0/16/32/48 px)
Determines interior crop offsets; swept result varies A-Avg by 1.3 points, supporting robustness rather than a cherry-picked optimum.
per-crop decode budget = 30 characters
Cap for per-crop greedy decoding. Contradicts the stated training cap L_train=25 (§3.1 vs §3.5.1), so crops decoding to 26–30 characters are technically out-of-distribution on the decoder axis; the PARSeq adapter uses a 25-character cap instead.
overlap-match minimum-similarity threshold = not stated in manuscript
§3.5.2 falls back to the geometric overlap estimate when the edit-distance similarity is below a threshold, but the threshold value is never reported, leaving a hidden knobs in the stitcher.
width-gating threshold = not stated numerically ('matched to training width')
§3.5.3 and §4.13 recommend applying chunk-decode-stitch only above a width threshold, but the concrete pixel value is not given; presumably 128 px but the reader must infer it.
fine-tuning two-stream weight and budget = 80/20 short-weighted; 5,000 iterations on weak base, 15,000 on strong base
Recommended operating point of the secondary fine-tuning recipe (§3.4, §4.5, §4.6); swept weights 50/50, 70/30, 80/20 and budgets 5k/15k/20k. Not part of the central training-free claim.

assumptions (4)

domain assumption LTB ground truth and exact-match protocol are fair enough to rank methods.
§4.1 and §4.10: the paper's own manual inspection found LTB typos (lmage, 6flfeb) that the model reads correctly but scores as failures; the benchmark is still used to compare against published SMTR/LISTER/FocalSVTR numbers.
domain assumption Published baseline numbers (Table 3 of Du et al. [10]) are directly comparable to the authors' evaluations.
§4.10 and §5: the headline comparison uses published rows, and the authors acknowledge one comparison mixes a full Union14M-L base against methods retrained on the filtered variant, and the PARSeq row uses a different corpus.
domain assumption The 36-character alphanumeric, case-insensitive, space/punctuation-stripped protocol defines the task.
§4.9: word boundaries, case, and punctuation are stripped from scoring, so the method's word accuracy is measured on a character-stream task, which is the standard LTB convention but bounds the practical claim about word-level reading.
standard math Extending the additive sinusoidal decoder table is a lossless buffer extension.
§3.3: recomputing the deterministic sinusoidal table at larger length leaves original positions numerically unchanged; used to argue the decoder axis fails in weights, not representation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Out-of-Length Scene Text Recognition: A Two-Axis Diagnosis and a Training-Free Fix." pith.science (2026). https://pith.science/paper/M3MW27MB

@misc{pith2026260723194,
  author       = {Pith},
  title        = {Pith review of: Out-of-Length Scene Text Recognition: A Two-Axis Diagnosis and a Training-Free Fix},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3MW27MB}},
  note         = {Machine review of arXiv:2607.23194}
}

read the original abstract

Scene Text Recognition (STR) models are trained almost exclusively on word crops of at most 25 characters, yet real deployments (signage, product labels, dense captions) require reading much longer text. This paper diagnoses that failure and then closes it. The diagnosis separates out-of-length failure into two simultaneously extrapolating axes (the encoder's width axis and the decoder's time axis) and shows that encoder width, not decoder length, is the dominant failure mode. Representation-side fixes bring only partial relief: training-free rotary rescalings recover at most 2-4 points of character error rate (CER), and a weighted fine-tuning recipe recovers 6-8 points while improving standard-benchmark accuracy, yet word accuracy on the Long Text Benchmark (LTB) stays near zero, because the residual gap lies in the decoding mechanism rather than the representation. We then close that gap at inference time, on an unmodified word-level checkpoint: the long image is sliced into overlapping crops at the model's training width, each decoded independently and in-distribution, and the reads stitched by geometry-anchored edit-distance alignment. This procedure reaches 42.79-43.05% bucket-average word accuracy on LTB across two base checkpoints, matching the published state of the art (41.57%) and beating it by 11-12 points on the hardest bucket, at wall-clock parity with plain decoding; applied unchanged to the public PARSeq checkpoint it reaches 47.11%. Once chunking is applied fine-tuning no longer helps: the decoding-side fix alone matches purpose-built architectures. We release the diagnosis harness and implementation.

Figures

Figures reproduced from arXiv: 2607.23194 by the authors.

**Figure 1.** chunk-decode-stitch: a training-free, inference-time pipeline that turns an unmodified wordlevel STR checkpoint fθ into a long-text reader. The wide image is aspect-preserving resized to the training height, sliced into overlapping crops at exactly the model’s training width, each crop decoded independently by the frozen greedy decoder (so every decode is in-distribution on both the encoder width axis and the decod… view at source ↗

**Figure 2.** Qualitative both-axis reads on LTB: full-convergence base (“Base”) vs. the recommended 80/ [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗

**Figure 3.** chunk-decode-stitch on real LTB hardest-bucket (L ≥ 56) examples, base checkpoint, stride 64. Each panel shows the aspect-preserved image strip (top) and the overlapping 128px crops actually fed to the frozen decoder, staggered downward with their independent in-distribution reads. (a) A sixcrop stitch producing an exact match on a multi-line document image, where every crop consistently reads the target line. (b) … view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Robustness of chunk-decode-stitch to its only two design choices, base checkpoint, LTB bothaxis A-Avg. Left: the crop-stride sweep shows a genuine plateau at ≥50% overlap (shaded), not a cherrypicked optimum, degrading only as overlap falls below it. Right: the crop-…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 2 canonical work pages

[10]

Y . Du, Z. Chen, C. Jia, X. Gao, Y .-G. Jiang, Out of length text recognition with sub-string matching, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 2025, pp. 2775–2783

2025
[1]

Jaderberg, K

M. Jaderberg, K. Simonyan, A. Vedaldi, A. Zisserman, Reading text in the wild with convolutional neural networks, IJCV 116 (2016) 1–20

2016
[2]

Gupta, A

A. Gupta, A. Vedaldi, A. Zisserman, Synthetic data for text localisation in natural images, in: CVPR, 2016, pp. 2315–2324

2016
[3]

Mishra, K

A. Mishra, K. Alahari, C. V . Jawahar, Scene text recognition using higher order language priors, in: BMVC, 2012

2012
[4]

K. Wang, B. Babenko, S. Belongie, End-to-end scene text recognition, in: ICCV , 2011, pp. 1457– 1464

2011
[5]

Karatzas, F

D. Karatzas, F. Shafait, S. Uchida, M. Iwamura, L. G. i Bigorda, S. R. Mestre, J. Mas, D. F. Mota, J. A. Almazan, L. P. de las Heras, ICDAR 2013 robust reading competition, in: ICDAR, 2013, pp. 1484–1493

2013
[6]

Karatzas, L

D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. Ghosh, A. Bagdanov, M. Iwamura, J. Matas, L. Neu- mann, V . R. Chandrasekhar, S. Lu, et al., ICDAR 2015 competition on robust reading, in: ICDAR, 2015, pp. 1156–1160

2015
[7]

T. Q. Phan, P. Shivakumara, S. Tian, C. L. Tan, Recognizing text with perspective distortion in natural scenes, in: ICCV , 2013, pp. 569–576

2013

Show all 47 references

[8]

Risnumawan, P

A. Risnumawan, P. Shivakumara, C. S. Chan, C. L. Tan, A robust arbitrary text detection system for natural scene images, Expert Systems with Applications 41 (2014) 8027–8048

2014
[9]

C. Cui, R. Wang, J. Si, C. Li, X. Wang, L. Peng, Lister: Neighbor decoding for length-insensitive scene text recognition, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 19541–19551

2023
[11]

J. Su, M. Ahmed, Y . Lu, S. Pan, W. Bo, Y . Liu, RoFormer: Enhanced transformer with rotary position embedding, Neurocomputing 568 (2024) 127063

2024
[12]

S. Chen, S. Wong, L. Chen, Y . Tian, Extending context window of large language models via posi- tional interpolation, arXiv preprint arXiv:2306.15595 (2023)

2023 arXiv
[13]

B. Peng, J. Quesnelle, H. Fan, E. Shippole, Yarn: Efficient context window extension of large lan- guage models, in: International Conference on Learning Representations (ICLR), 2024

2024
[14]

Raisi, 2d rotary position embedding for scene text recognition with transformers, arXiv preprint arXiv:2607.13458 (2026).arXiv:2607.13458

Z. Raisi, 2d rotary position embedding for scene text recognition with transformers, arXiv preprint arXiv:2607.13458 (2026).arXiv:2607.13458

2026 arXiv
[15]

Jiang, J

Q. Jiang, J. Wang, D. Peng, C. Liu, L. Jin, Revisiting scene text recognition: A data perspective, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 20543–20554

2023
[16]

Atienza, Vision transformer for fast and efficient scene text recognition, in: ICDAR, 2021, pp

R. Atienza, Vision transformer for fast and efficient scene text recognition, in: ICDAR, 2021, pp. 319–334

2021
[17]

S. Fang, H. Xie, Y . Wang, Z. Sun, J. Gao, Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition, in: CVPR, 2021, pp. 7098–7107

2021
[18]

D. B. Bautista, R. Atienza, Scene text recognition with permuted autoregressive sequence models, in: ECCV , 2022, pp. 178–196

2022
[19]

M. Yang, Y . Guo, Z. Chen, J. Guo, Reading and writing: Discriminative and generative modeling for self-supervised text recognition, in: CVPR, 2022, pp. 7329–7338

2022
[20]

Afkari-Fahandari, E

A. Afkari-Fahandari, E. Shabaninia, F. Asadi-Zeydabadi, H. Nezamabadi-Pour, A comprehensive survey of transformers in text recognition: Techniques, challenges, and future directions, ACM Com- puting Surveys 58 (2025). doi:10.1145/3771273

2025 doi
[21]

Kadha, B

V . Kadha, B. B. Duddeti, K. Srinadh, S. K. Buddepu, L. Janjanam, K. Medhi, From pixels to text: A deep learning survey of scene text detection and recognition, Computers and Electrical Engineering 135 (2026) 111139. doi:10.1016/j.compeleceng.2026.111139

2026
[22]

B. Shi, X. Bai, C. Yao, An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition, IEEE TPAMI 39 (2016) 2298–2304

2016
[23]

B. Shi, M. Yang, X. Wang, P. Lyu, C. Yao, X. Bai, ASTER: An attentional scene text recognizer with flexible rectification, IEEE TPAMI 41 (2018) 2035–2048

2018
[24]

F. Zhan, C. Lu, ESIR: End-to-end scene text recognition via iterative image rectification, in: CVPR, 2019, pp. 2059–2068

2019
[25]

Y . Baek, G. Lee, D. Yun, H. Lee, What is wrong with scene text recognition model comparisons? dataset and model analysis, in: ICCV , 2021, pp. 4715–4723. 34

2021
[26]

Y . Du, Z. Chen, Y . Su, C. Jia, Y . Jiang, Context perception parallel decoder for scene text recognition, IEEE Transactions on Pattern Analysis and Machine Intelligence (2025) 4668–4683

2025
[27]

B. Zhou, A. Zhu, J. Yang, MambaSTR: Scene text recognition with masked state space model, in: Advanced Intelligent Computing Technology and Applications (ICIC), 2025

2025
[28]

S. Ali, T. Gao, Z. Zhu, Z. Hussain, M. Ahmed, Mamba-str: Efficient context-aware scene text recog- nition framework with selective state space modeling, Information Sciences (2026) 123242

2026
[29]

Tian, K.-X

S. Tian, K.-X. Zhu, H.-B. Qin, C. Yang, Dynamic receptive field adaptation for scene text recognition, Pattern Recognition Letters 178 (2024) 55–61. doi:10.1016/j.patrec.2023.12.005

2024 doi
[30]

Y . Du, Z. Chen, H. Xie, C. Jia, Y .-G. Jiang, Svtrv2: Ctc beats encoder-decoder models in scene text recognition, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 20147–20156

2025
[31]

D. H. Diaz, S. Qin, R. Ingle, Y . Fujii, A. Bissacco, Rethinking text line recognition models, arXiv preprint arXiv:2104.07787 (2021)

2021 arXiv
[32]

Nguyen, N

D. Nguyen, N. Tran, H. Le, Improving long handwritten text line recognition with convolutional multi-way associative memory, arXiv preprint arXiv:1911.01577 (2019)

1911 arXiv
[33]

Garrido-Munoz, J

C. Garrido-Munoz, J. Calvo-Zaragoza, On the generalization of handwritten text recognition models, in: CVPR, 2025

2025
[34]

Press, N

O. Press, N. A. Smith, M. Lewis, Train short, test long: Attention with linear biases enables input length extrapolation, in: ICLR, 2022

2022
[35]

L. Zhao, X. Feng, X. Feng, W. Zhong, D. Xu, Q. Yang, H. Liu, B. Qin, T. Liu, Length extrapolation of Transformers: A survey from the perspective of positional encoding, Findings of the Association for Computational Linguistics: EMNLP 2024 (2024)

2024
[36]

Barbero, A

F. Barbero, A. Vitvitskyi, C. Perivolaropoulos, R. Pascanu, P. Veli ˇckovi´c, Round and round we go! What makes rotary positional encodings useful?, in: International Conference on Learning Representations (ICLR), 2025

2025
[37]

Veisi, D

A. Veisi, D. Fartoot, H. Amirzadeh, Context-aware rotary position embedding, arXiv preprint arXiv:2507.23083 (2025)

2025 arXiv
[38]

Dosovitskiy, L

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al., An image is worth 16x16 words: Transformers for image recognition at scale, in: ICLR, 2021

2021
[39]

B. Heo, S. Yun, D. Han, S. Chun, C. Choi, Y . Oh, J. Kim, Rotary position embedding for vision transformer, in: ECCV , 2024, pp. 336–352

2024
[40]

H. Liu, S. Ren, T. Zhu, P. Wang, C. Xie, A. Yuille, Z. Zheng, F. Wang, Spiral RoPE: Rotate your rotary positional embeddings in the 2D plane, arXiv preprint arXiv:2602.03227 (2026)

2026
[41]

Raisi, M

Z. Raisi, M. A. Naiel, P. Fieguth, S. Wardell, J. Zelek, 2D positional embedding-based transformer for scene text recognition, Journal of Computational Vision and Imaging Systems 6 (2020) 1–4

2020
[42]

Raisi, M

Z. Raisi, M. A. Naiel, G. Younes, S. Wardell, J. S. Zelek, 2LSPE: 2D learnable sinusoidal positional encoding using transformer for scene text recognition, in: 18th Conference on Robots and Vision (CRV), IEEE, 2021, pp. 119–126

2021
[43]

Raisi, V

Z. Raisi, V . M. N. Had, A transformer-based approach with contextual position encoding for robust Persian text recognition in the wild, Journal of Artificial Intelligence and Data Mining (JAIDM) 12 (2024) 455–464

2024
[44]

X. Chen, S. Zhou, M. Huang, J. Feng, Y . Xiong, K. Zhou, B. Yang, Y . Zhang, H. Bao, S. Peng, C. Li, F. Shi, A 2d semantic-aware position encoding for vision transformers, arXiv preprint arXiv:2505.09466 (2025)

2025 arXiv
[45]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, in: NeurIPS, 2017, pp. 5998–6008

2017
[46]

K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: CVPR, 2016, pp. 770–778

2016
[47]

gain erased at the long tail

Y . Du, Z. Chen, C. Jia, X. Yin, T. Zheng, C. Li, Y . Du, Y .-G. Jiang, SVTR: Scene text recognition with a single visual model, in: Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI), 2022, pp. 884–890. 35 30 40 50 60 70 80 90 10...

2022

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [10]

Y . Du, Z. Chen, C. Jia, X. Gao, Y .-G. Jiang, Out of length text recognition with sub-string matching, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 2025, pp. 2775–2783

2025

[2] [1]

Jaderberg, K

M. Jaderberg, K. Simonyan, A. Vedaldi, A. Zisserman, Reading text in the wild with convolutional neural networks, IJCV 116 (2016) 1–20

2016

[3] [2]

Gupta, A

A. Gupta, A. Vedaldi, A. Zisserman, Synthetic data for text localisation in natural images, in: CVPR, 2016, pp. 2315–2324

2016

[4] [3]

Mishra, K

A. Mishra, K. Alahari, C. V . Jawahar, Scene text recognition using higher order language priors, in: BMVC, 2012

2012

[5] [4]

K. Wang, B. Babenko, S. Belongie, End-to-end scene text recognition, in: ICCV , 2011, pp. 1457– 1464

2011

[6] [5]

Karatzas, F

D. Karatzas, F. Shafait, S. Uchida, M. Iwamura, L. G. i Bigorda, S. R. Mestre, J. Mas, D. F. Mota, J. A. Almazan, L. P. de las Heras, ICDAR 2013 robust reading competition, in: ICDAR, 2013, pp. 1484–1493

2013

[7] [6]

Karatzas, L

D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. Ghosh, A. Bagdanov, M. Iwamura, J. Matas, L. Neu- mann, V . R. Chandrasekhar, S. Lu, et al., ICDAR 2015 competition on robust reading, in: ICDAR, 2015, pp. 1156–1160

2015

[8] [7]

T. Q. Phan, P. Shivakumara, S. Tian, C. L. Tan, Recognizing text with perspective distortion in natural scenes, in: ICCV , 2013, pp. 569–576

2013