2D Rotary Position Embedding for Scene Text Recognition with Transformers

Zobeir Raisi

REVIEW 3 major objections 5 minor 1 cited by

2D Rotary Position Embedding for Scene Text Recognition with Transformers

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Rotating attention by row and column offsets helps Transformer scene-text recognizers read curved and perspective-warped words, lifting average accuracy to 90.4% with no added parameters.

desk verdict A useful 2D-RoPE adaptation for STR with a good diagnostic protocol, but the headline curved-text claim is contradicted by the controlled data and the design ablations are internally inconsistent. read the letter →

arxiv 2607.13458 v1 pith:4VN7NRB3 submitted 2026-07-15 cs.CV

Zobeir Raisi This is my paper · ORCID

classification cs.CV

keywords SceneTextRecognitionRotaryPositionEmbedding2DpositionalencodingTransformerirregularrelativecross-attentionattentionvisualization

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

This paper sets out to show that the way positions are encoded inside a Transformer matters for reading text in natural scenes, and that the standard 1D positional encodings inherited from language models are the wrong tool. It adapts 2D rotary position embedding (2D-RoPE) to scene text recognition by rotating query and key vectors along separate row and column axes, allocating more dimensions to the row axis to match text's wide aspect ratio, and extending the rotation into decoder cross-attention. The central claim is that attention scores then depend only on the relative 2D offset between tokens, so the encoder can follow the reading order of curved, rotated, or perspective-distorted text without any rectification module. On six benchmarks the model reaches the best average accuracy (90.4%), with the largest gains on exactly the irregular layouts — curved CUTE80 (+3.8) and perspective-distorted SVTP (+1.5) — and a controlled ablation pair isolates the positional encoding as the source of the difference. The result matters because the module is plug-and-play and adds only two scalar parameters, offering a near-free accuracy improvement for any Transformer-based STR system.

What carries the argument

The central object is the axial 2D Rotary Position Embedding (2D-RoPE): each token's feature vector is split into two halves, the first rotated by the row index and the second by the column index using block-diagonal rotation matrices with frequency bases θ_i = 10000^{-2(i-1)/d}. Theorem 1 proves the rotation makes the attention dot product a function of the relative offsets (Δh, Δw) only, so global translation of the character grid does not change attention. Two modifications carry the argument: an anisotropic dimension split (384 row : 128 column for a 512-dimension vector, matched to text's width-to-height ratio) with learnable per-axis frequency scales β_h, β_w, and an extension into dec

What would settle it

Evaluate the same ablation pair on a controlled set of text crops rotated well beyond the ±15° training augmentation — e.g., 45° slants or vertically stacked characters spanning multiple rows of the 8×32 grid. The claim predicts 2D-RoPE should retain a positive net-win count over 1D RoPE; a reversal, or attention maps that smear across rows, would show the axial relative-position mechanism does not track reading order on such layouts.

Watch

Extended reading notes

Core claim

The paper claims that 2D-RoPE-STR, a Transformer scene-text recognizer rotating queries and keys by row and column indices, reaches the best average accuracy (90.4%) on six benchmarks, ahead of the prior best (89.5%). The load-bearing identity is Theorem 1: after the axial rotation, attention depends only on the relative offset (Δh, Δw), turning absolute grid positions into relative 2D geometry. The author argues this is why irregular text is read correctly — attention locks onto strokes and follows reading order on curved or slanted words, while a 1D-sinusoidal baseline confuses characters. Two adaptations distinguish this from vision-only 2D-RoPE: an uneven 384:128 row/column split matched

Load-bearing premise

The method assumes the 8×32 CNN feature grid preserves a layout in which the row axis spans a single line of characters, so relative row/column offsets track character displacement along reading order; if curved or slanted text wraps characters across grid rows, the axial rotation stops following reading order — a limitation the paper concedes in §5.

Editorial extensions

If this is right

Any Transformer-based STR recognizer can replace its positional encoding with 2D-RoPE and expect gains on curved and perspective-distorted text, with no backbone changes required.
Gains concentrate where reading order departs from a horizontal line: net-win counts favour 2D-RoPE by +26 on IC15, +12 on SVTP, and +7 on SVT, while regular-text sets stay within noise.
Extending the rotary coupling to decoder cross-attention adds a measurable improvement (average 87.19→87.82 under the reduced schedule), showing the decoder also benefits from 2D layout awareness.
With only two learnable scale parameters, 2D-RoPE beats learned additive 2D encodings (which cost roughly 49K parameters) in both accuracy and parameter efficiency.
The diagnostic protocol — ablation pair, per-image disagreement analysis, attention maps — offers future positional-encoding studies a template for isolating the positional module from architecture and training effects.

Reading between the lines

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

Because the mechanism is translation-invariant in 2D, combining 2D-RoPE with explicit rectification (e.g., a spatial transformer) is a testable extension; the paper's own view is that the two are complementary, so a fused model should outperform either alone on curved text.
The axial split is the load-bearing constraint: a benchmark of steeply slanted or vertically stacked text, where characters wrap across multiple rows of the 8×32 grid, should reverse the advantage toward additive 2D encodings or force an adaptive dimension split — a direct test of the limitation flagged in §5.
If multi-directional rotary variants (as the author suggests) replace the axis-aligned split, the prediction is they mainly recover the curved-benchmark deficit (CUTE80 net −1 in Table 4) while leaving regular-text performance unchanged.
The near-zero parameter cost suggests the relative ordering should persist at higher resolutions or with larger backbones, where learned position tables scale poorly; scaling the same ablation pair to a 64×256 input grid would check this.

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

Quick take: this paper has a genuinely useful idea — carrying 2D rotary positions into an STR encoder-decoder with a row-heavy split and cross-attention key rotation — and the controlled disagreement analysis is a good diagnostic habit. But the evidence as reported doesn't support the central claim that curved text is a key beneficiary, and the design ablations in Table 3 are internally inconsistent. It deserves a serious referee, but the authors need to fix the numbers and rerun before the claims can be trusted.

What is new: applying axial 2D-RoPE to scene text recognition with an anisotropic split (384:128 instead of 256:256) and rotating cross-attention keys is a sensible adaptation. The math is standard, but the paper gives a clear restatement of the relative-position property, and the diagnostic protocol — the controlled ablation pair, the image-level disagreement count, and attention visualizations — is a reasonable effort to isolate where a positional encoding matters.

What works: the controlled pair in Table 4 shows a net +37/7672 images, with gains on IC15 and SVTP, which is consistent with the claim that relative 2D position helps some irregular layouts. The limitations section honestly flags the axial decomposition's weakness on oblique text.

Where it softens: the abstract's headline numbers (+3.8 CUTE80, +1.5 SVTP over MATRN) are uncontrolled cross-model comparisons. The paper's own controlled ablation shows CUTE80 net -1, so the curved-text emphasis is not supported. Worse, Table 3's blocks are internally inconsistent: the 'def.' rows vary from 86.31 to 87.82 Avg across blocks, meaning the default configuration was not a fixed baseline. That undermines the within-block comparisons. §4.4.1 quotes SVTP values that don't appear in Table 2, so the perspective-text ablation claim is unverifiable. And two learnable scalars aren't 'parameter-free', though 'essentially' does some work there. No error bars or multiple seeds are reported, which matters for a net gain of ~0.5%.

This paper is for researchers working on positional encodings in STR or vision transformers generally. It's a useful case study, but only after the empirical inconsistencies are addressed. My recommendation: send it to peer review, but with a clear request for corrected tables, reruns with seeds, and a reframed abstract that matches the controlled evidence.

Referee Report

3 major / 5 minor

Summary. The paper proposes 2D-RoPE-STR, a rotary position embedding for scene text recognition that applies axial rotations to query and key vectors along both the row and column axes of the CNN feature grid, extends the rotary coupling to encoder–decoder cross-attention, and allocates rotary dimensions unevenly between rows and columns to match text anisotropy. The authors claim that this encoding makes attention depend only on relative 2D offsets, and that it yields accuracy gains on irregular text, especially curved and perspective-distorted benchmarks, at essentially zero parameter cost, with a controlled ablation pair and diagnostic analysis to support the claim.

Significance. The mathematical core of the paper, Eqs. (5)–(7) and Theorem 1, is correct: rotating disjoint channel groups by row and column indices gives an attention score that depends only on the relative offsets (Δh, Δw). This is a clean, well-known extension of RoPE to 2D, and the specific adaptations to STR—anisotropic dimension split and cross-attention application—are plausible and plug-and-play. The controlled ablation pair in Table 4 is a genuinely good design: it isolates the positional encoding while holding the backbone, decoder, and schedule fixed, and it honestly reports a small net effect (+37/7,672 images). The paper also provides code, checkpoints, and an image-level disagreement analysis, which are strengths. However, the empirical support for the headline claim about curved text is internally inconsistent and weaker than the abstract suggests. The controlled evidence shows no gain on CUTE80, while the cross-model comparison to MATRN, which is not controlled, is used to claim +3.8 on CUTE80. Several load-bearing ablations report numbers that conflict across tables. These issues must be resolved before the central claim can be accepted.

major comments (3)

[§4.4.2, Table 3] The design ablations are internally inconsistent. The default configuration is supposed to be common across the four blocks, yet the reported numbers for the default differ: in Table 3(a) the learnable-β default has CUTE80=86.11 and Avg=86.31; in Table 3(b) the 384:128 default has CUTE80=86.11 and Avg=86.92; in Table 3(c) the 6-layer default has CUTE80=86.81 and Avg=87.18; and in Table 3(d) the 2D-RoPE default has CUTE80=88.54 and Avg=87.82. Since the same default configuration appears with different performance in different rows, the claim that each block varies exactly one factor around a common default cannot be checked and the comparative conclusions—e.g., that 384:128 is best or that cross-attention 2D-RoPE improves over none—are not reliable.
[§4.4.1, Table 2] The text states that removing positional encoding gives the steepest degradation on perspective text, citing SVTP values of 78.8% vs. 81.2%, and claims a consistent ordering across six benchmarks. However, Table 2 reports only IIIT5K, SVT, CUTE80, and the average; SVTP and the other two benchmarks are absent. The six-benchmark average cannot be verified, and the specific SVTP numbers quoted in §4.4.1 do not appear anywhere in the table. This is a load-bearing inconsistency because the paper's central empirical claim is that gains concentrate on perspective-distorted text.
[§4.3, Table 1 vs. §4.5, Table 4] The abstract and §1 emphasize large gains on curved text (+3.8 CUTE80) and perspective text (+1.5 SVTP) over MATRN. These comparisons are not controlled: MATRN and 2D-RoPE-STR differ in architecture, training, and possibly data. The controlled ablation pair in Table 4 shows a net gain of +37 images out of 7,672 (~0.5%), with CUTE80 at net −1 and SVTP at +12. No multiple-seed runs or confidence intervals are reported, so even the +37 could be run-to-run variation. The paper should either report the controlled results with appropriate uncertainty or substantially soften the claim that curved text is a key beneficiary; the current presentation overstates the evidence.

minor comments (5)

[§3.3.3, §4.4.1] The phrase 'essentially parameter-free' is overstated: β_h and β_w are learned parameters, even though only two scalars. Moreover, Table 3(a) shows the learnable-β default has a lower average (86.31) than the fixed bases (86.71 for θ=10000 and 86.66 for mixed), so the 'default' is not justified by the reported ablations.
[§5, Limitations] The Limitations section concedes that the axial decomposition is 'less expressive for oblique or diagonal arrangements.' This directly explains why the controlled CUTE80 result in Table 4 is net negative. The abstract and §1 should be reconciled with this acknowledged limitation rather than presenting curved text as a primary success story.
[Table 2] Since the ablation protocol is intended to evaluate all six benchmarks, Table 2 should report all six columns, or the text should explicitly state why only three are shown.
[Table 3] The table header lists 'Configuration' followed by CUTE80 and Avg, but the rows are not self-explanatory. Please add a column indicating the default configuration explicitly and mark all numbers that correspond to the default in each block.
[§4.5, Fig. 4] The attention visualization is qualitative. Since the paper introduces a 'diagnostic protocol,' it would be more convincing to provide a quantitative comparison of attention concentration or a metric that contrasts the 2D-RoPE and baseline attention maps.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the relative-position theorem is an algebraic identity, and the empirical claims rest on external benchmarks and controlled ablations.

full rationale

The paper's derivation chain does not reduce to its own inputs. Theorem 1 (Eq. 7) is a direct algebraic consequence of the block-diagonal rotation definition of 2D-RoPE; it restates the standard RoPE relative-position identity for two spatial axes and is not fitted to any outcome. The central claim—that 2D relative position helps on irregular text—is tested against six standard public benchmarks and a controlled ablation pair in which only the positional encoding differs (Tables 1-4). These are external empirical checks, not constructions from the method's own parameters. The self-citations [33-35] are used only in Related Work to frame prior additive 2D encodings and identify the gap; they are not cited as evidence for the new method's performance, and no uniqueness theorem or prior result is invoked to force the design. The learnable scales beta_h, beta_w are trained parameters, not renamed predictions; the 384:128 split is a design choice selected from ablation tables, which is a test-set-selection concern rather than a definitional loop. The skeptic's points about uncontrolled cross-model comparisons, missing error bars, and the CUTE80 discrepancy in the controlled ablation are validity/correctness weaknesses, not circularity, because the reported numbers are not equivalent by construction to the method's assumptions. The Limitations section's concession that the axial decomposition is less expressive for oblique or diagonal arrangements is an honest boundary on the claim, not evidence that the result was assumed into existence.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or model entities; 2D-RoPE is an operation on existing features and β_h,β_w are two scalar parameters. The load-bearing free choices are the two scalars and the hand-chosen 384:128 split, both selected post hoc on the same test benchmarks. The underlying 'axioms' are standard rotation math plus a domain assumption about grid layout alignment with reading order, plus the experiment-design premise that Table 3 blocks vary one factor — which the data contradict.

free parameters (2)

learnable per-axis frequency scale factors β_h, β_w = initialized to 1.0; learned values not reported
Introduced in §3.3.3 (Eq. 11) as the default configuration; §4.4.1 confirms 2 trained scalars total. In Table 3(a) this default is the worst of four options on CUTE80 (86.11 vs 88.19 for fixed θ=500), yet it is carried to the main model.
row/column dimension split (d_h:d_w = 384:128) = 384:128 over d=512
Selected in §4.4.2 Table 3(b) as best on the same test benchmarks; justified heuristically by text aspect ratio (wider than tall). It is a fixed choice, not adaptive to input dimensions.

assumptions (4)

standard math RoPE rotation matrices are orthogonal and block-diagonal, giving ⟨R_m q, R_n k⟩ = ⟨q, R_{n−m} k⟩ (Eq. 3); Theorem 1 applies this per-axis (Eq. 7).
Proven in Su et al. [40] and re-derived in the paper's Theorem 1; standard linear algebra of rotation matrices.
domain assumption The flattened 8×32 feature grid from a stride-4 ResNet-50 preserves a meaningful 2D character layout: the row axis tracks vertical position and the column axis horizontal position, so relative offsets correspond to reading order.
Section 3.3.1: 'each token retaining its 2D spatial coordinate (h, w)'. This is the premise for why 2D relative position should help irregular text; the paper's own Limitations (§5) concede axial decomposition is weak for oblique/diagonal arrangements.
ad hoc to paper Decoder cross-attention queries are treated as position-free while encoder keys carry 2D-RoPE; the relative-position signal between query and key is therefore partial (key-side only).
Section 3.3.4 states decoder queries 'do not have 2D spatial positions' and applies RoPE only to keys; the asymmetry is a design choice validated empirically in Table 3(d), not derived.
ad hoc to paper Each block of Table 3 varies exactly one factor around a common default configuration.
Section 4.4.2 asserts 'Each block varies one factor', but the reported 'def.' rows disagree across blocks (Avg 86.31–87.82; CUTE80 86.11–88.54), so this premise fails in the reported data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of 2D Rotary Position Embedding for Scene Text Recognition with Transformers." pith.science (2026). https://pith.science/paper/4VN7NRB3

@misc{pith2026260713458,
  author       = {Pith},
  title        = {Pith review of: 2D Rotary Position Embedding for Scene Text Recognition with Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4VN7NRB3}},
  note         = {Machine review of arXiv:2607.13458}
}

read the original abstract

Scene Text Recognition (STR) remains challenging due to the diversity of text appearances, including curvature, rotation, and perspective distortion. Recent Transformer-based approaches perform well but usually rely on one-dimensional positional encodings that ignore the 2D spatial structure of text images. Axial 2D extensions of Rotary Position Embedding (RoPE) exist for vision Transformers, but they assume roughly square, isotropic image content and apply the rotation only within encoder self-attention. Scene text violates both assumptions: crops are markedly anisotropic, and STR models are encoder-decoder, so the decoder must relate its queries to the encoder's 2D layout through cross-attention. We introduce 2D-RoPE-STR, which adapts axial 2D-RoPE to this setting through (1) an anisotropic row/column dimension allocation matched to the aspect ratio of text, and (2) an extension of the rotary coupling into encoder-decoder cross-attention, letting autoregressive decoding steps attend to encoder tokens by their 2D layout, a setting not addressed by prior encoder-only formulations. Both changes are essentially parameter-free and require no architectural redesign beyond the positional-encoding module. We further introduce a diagnostic protocol (a controlled ablation pair isolating only the positional encoding, an image-level net-win disagreement analysis, and encoder attention visualization) that identifies where and why relative 2D position helps: curved, rotated, and perspective-distorted layouts where reading order departs from a straight horizontal line. On six standard benchmarks (IIIT5K, SVT, ICDAR 2013, ICDAR 2015, CUTE80, SVTP), gains concentrate on exactly these irregular layouts, with ablations isolating each design choice against 1D RoPE and 2D sinusoidal and learnable alternatives.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Out-of-Length Scene Text Recognition: A Two-Axis Diagnosis and a Training-Free Fix
cs.CV 2026-07 conditional novelty 5.0 of 10

A training-free chunk-and-stitch procedure turns ordinary word-level scene text recognizers into near-state-of-the-art long-text readers, after the paper shows the encoder-width axis, not decoder length, dominates the...

Reference graph

Works this paper leans on

48 extracted references · cited by 1 Pith paper

[1]

ACM Computing Surveys 58(5)

Afkari-Fahandari A, Shabaninia E, Asadi-Zeydabadi F, et al (2025) A comprehen- sive survey of transformers in text recognition: Techniques, challenges, and future directions. ACM Computing Surveys 58(5). https://doi.org/10.1145/3771273

doi:10.1145/3771273 2025
[2]

Information Sciences p 123242

Ali S, Gao T, Zhu Z, et al (2026) Mamba-str: Efficient context-aware scene text recognition framework with selective state space modeling. Information Sciences p 123242

2026
[3]

In: ICDAR, pp 319–334

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

2021
[4]

In: ICCV, pp 4715–4723

Baek Y, Lee G, Yun D, et al (2021) What is wrong with scene text recognition model comparisons? dataset and model analysis. In: ICCV, pp 4715–4723

2021
[5]

Barbero F, Vitvitskyi A, Perivolaropoulos C, et al (2024) Round and round we go! What makes rotary positional encodings useful? arXiv preprint arXiv:241006205

2024
[6]

In: ECCV, pp 178–196

Bautista DB, Atienza R (2022) Scene text recognition with permuted autoregres- sive sequence models. In: ECCV, pp 178–196

2022
[7]

In: AAAI, pp 1430–1437

Bhunia AK, Bhunia AK, Sain A, et al (2021) Joint visual-linguistic representation for text recognition. In: AAAI, pp 1430–1437

2021
[8]

In: ECCV, pp 213–229

Carion N, Massa F, Synnaeve G, et al (2020) End-to-end object detection with transformers. In: ECCV, pp 213–229

2020

Show all 48 references

[9]

In: ICCV, pp 5070–5078

Cheng Z, Bai F, Xu Y, et al (2017) Focusing attention: Towards accurate text recognition in natural images. In: ICCV, pp 5070–5078

2017
[10]

JMLR 24(240):1–113

Chowdhery A, Narang S, Devlin J, et al (2023) PaLM: Scaling language modeling with pathways. JMLR 24(240):1–113

2023
[11]

In: NeurIPS

Chu X, Tian Z, Wang Y, et al (2021) Conditional positional encodings for vision transformers. In: NeurIPS

2021
[12]

In: NAACL-HLT, pp 4171–4186

Devlin J, Chang MW, Lee K, et al (2019) BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL-HLT, pp 4171–4186

2019
[13]

In: ICLR 14

Dosovitskiy A, Beyer L, Kolesnikov A, et al (2021) An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR 14

2021
[14]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp 20147–20156

Du Y, Chen Z, Xie H, et al (2025) Svtrv2: Ctc beats encoder-decoder mod- els in scene text recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp 20147–20156

2025
[15]

In: CVPR, pp 7098– 7107

Fang S, Xie H, Wang Y, et al (2021) Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition. In: CVPR, pp 7098– 7107

2021
[16]

In: ICML, pp 369–376

Graves A, Fern´ andez S, Gomez F, et al (2006) Connectionist temporal classifi- cation: Labelling unsegmented sequence data with recurrent neural networks. In: ICML, pp 369–376

2006
[17]

In: CVPR, pp 2315–2324

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

2016
[18]

In: CVPR, pp 770–778

He K, Zhang X, Ren S, et al (2016) Deep residual learning for image recognition. In: CVPR, pp 770–778

2016
[19]

In: ECCV, pp 336–352

Heo B, Yun S, Han D, et al (2024) Rotary position embedding for vision transformer. In: ECCV, pp 336–352

2024
[20]

IJCV 116(1):1–20

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

2016
[21]

Computers and Electrical Engineering 135:111139

Kadha V, Duddeti BB, Srinadh K, et al (2026) From pixels to text: A deep learning survey of scene text detection and recognition. Computers and Electrical Engineering 135:111139. https://doi.org/10.1016/j.compeleceng.2026.111139

2026
[22]

In: ICDAR, pp 1484–1493

Karatzas D, Shafait F, Uchida S, et al (2013) ICDAR 2013 robust reading competition. In: ICDAR, pp 1484–1493

2013
[23]

In: ICDAR, pp 1156–1160

Karatzas D, Gomez-Bigorda L, Nicolaou A, et al (2015) ICDAR 2015 competition on robust reading. In: ICDAR, pp 1156–1160

2015
[24]

In: ICLR

Kingma DP, Ba J (2015) Adam: A method for stochastic optimization. In: ICLR

2015
[25]

In: CVPR, pp 2231–2239

Lee CY, Osindero S (2016) Recursive recurrent nets with attention modeling for OCR in the wild. In: CVPR, pp 2231–2239

2016
[26]

In: AAAI, pp 8610–8617

Li H, Wang P, Shen C, et al (2019) Show, attend and read: A simple and strong baseline for irregular text recognition. In: AAAI, pp 8610–8617

2019
[27]

arXiv preprint arXiv:260203227

Liu H, Ren S, Zhu T, et al (2026) Spiral RoPE: Rotate your rotary positional embeddings in the 2D plane. arXiv preprint arXiv:260203227

2026
[28]

In: BMVC 15

Mishra A, Alahari K, Jawahar CV (2012) Scene text recognition using higher order language priors. In: BMVC 15

2012
[29]

In: W ACV, pp 824–833

Nuriel O, Benny S, Aides A (2023) MATRN: Multimodal-aware text recognition network. In: W ACV, pp 824–833

2023
[30]

In: ICCV, pp 569–576

Phan TQ, Shivakumara P, Tian S, et al (2013) Recognizing text with perspective distortion in natural scenes. In: ICCV, pp 569–576

2013
[31]

In: ICLR

Press O, Smith NA, Lewis M (2022) Train short, test long: Attention with linear biases enables input length extrapolation. In: ICLR

2022
[32]

JMLR 21(140):1–67

Raffel C, Shazeer N, Roberts A, et al (2020) Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR 21(140):1–67

2020
[33]

Journal of Artificial Intelligence and Data Mining (JAIDM) 12(3):455–464

Raisi Z, Had VMN (2024) 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(3):455–464

2024
[34]

Journal of Computational Vision and Imaging Systems 6(1):1–4

Raisi Z, Naiel MA, Fieguth P, et al (2020) 2D positional embedding-based trans- former for scene text recognition. Journal of Computational Vision and Imaging Systems 6(1):1–4

2020
[35]

In: 18th Conference on Robots and Vision (CR V), IEEE, pp 119–126

Raisi Z, Naiel MA, Younes G, et al (2021) 2LSPE: 2D learnable sinusoidal posi- tional encoding using transformer for scene text recognition. In: 18th Conference on Robots and Vision (CR V), IEEE, pp 119–126

2021
[36]

Expert Systems with Applications 41(18):8027–8048

Risnumawan A, Shivakumara P, Chan CS, et al (2014) A robust arbitrary text detection system for natural scene images. Expert Systems with Applications 41(18):8027–8048

2014
[37]

In: NAACL-HLT, pp 464–468

Shaw P, Uszkoreit J, Vaswani A (2018) Self-attention with relative position representations. In: NAACL-HLT, pp 464–468

2018
[38]

IEEE TPAMI 39(11):2298–2304

Shi B, Bai X, Yao C (2016) An end-to-end trainable neural network for image- based sequence recognition and its application to scene text recognition. IEEE TPAMI 39(11):2298–2304

2016
[39]

IEEE TPAMI 41(9):2035–2048

Shi B, Yang M, Wang X, et al (2018) ASTER: An attentional scene text recognizer with flexible rectification. IEEE TPAMI 41(9):2035–2048

2018
[40]

Neurocomputing 568:127063

Su J, Ahmed M, Lu Y, et al (2024) RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing 568:127063

2024
[41]

Pattern Recognition Letters 178:55–61

Tian S, Zhu KX, Qin HB, et al (2024) Dynamic receptive field adaptation for scene text recognition. Pattern Recognition Letters 178:55–61. https://doi.org/ 10.1016/j.patrec.2023.12.005

2024 doi
[42]

arXiv preprint arXiv:230213971 16

Touvron H, Lavril T, Izacard G, et al (2023) LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:230213971 16

2023
[43]

In: NeurIPS, pp 5998–6008

Vaswani A, Shazeer N, Parmar N, et al (2017) Attention is all you need. In: NeurIPS, pp 5998–6008

2017
[44]

In: ICCV, pp 1457–1464

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

2011
[45]

In: ECCV, pp 339–355

Wang P, Da C, Yao C (2022) Multi-granularity prediction for scene text recognition. In: ECCV, pp 339–355

2022
[46]

In: ICCV, pp 9147–9156

Yang M, Liu Y, Jiang Z, et al (2019) Symmetry-constrained rectification network for scene text recognition. In: ICCV, pp 9147–9156

2019
[47]

In: CVPR, pp 7329–7338

Yang M, Guo Y, Chen Z, et al (2022) Reading and writing: Discriminative and generative modeling for self-supervised text recognition. In: CVPR, pp 7329–7338

2022
[48]

In: CVPR, pp 2059–2068 17

Zhan F, Lu C (2019) ESIR: End-to-end scene text recognition via iterative image rectification. In: CVPR, pp 2059–2068 17

2019

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

ACM Computing Surveys 58(5)

Afkari-Fahandari A, Shabaninia E, Asadi-Zeydabadi F, et al (2025) A comprehen- sive survey of transformers in text recognition: Techniques, challenges, and future directions. ACM Computing Surveys 58(5). https://doi.org/10.1145/3771273

doi:10.1145/3771273 2025

[2] [2]

Information Sciences p 123242

Ali S, Gao T, Zhu Z, et al (2026) Mamba-str: Efficient context-aware scene text recognition framework with selective state space modeling. Information Sciences p 123242

2026

[3] [3]

In: ICDAR, pp 319–334

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

2021

[4] [4]

In: ICCV, pp 4715–4723

Baek Y, Lee G, Yun D, et al (2021) What is wrong with scene text recognition model comparisons? dataset and model analysis. In: ICCV, pp 4715–4723

2021

[5] [5]

Barbero F, Vitvitskyi A, Perivolaropoulos C, et al (2024) Round and round we go! What makes rotary positional encodings useful? arXiv preprint arXiv:241006205

2024

[6] [6]

In: ECCV, pp 178–196

Bautista DB, Atienza R (2022) Scene text recognition with permuted autoregres- sive sequence models. In: ECCV, pp 178–196

2022

[7] [7]

In: AAAI, pp 1430–1437

Bhunia AK, Bhunia AK, Sain A, et al (2021) Joint visual-linguistic representation for text recognition. In: AAAI, pp 1430–1437

2021

[8] [8]

In: ECCV, pp 213–229

Carion N, Massa F, Synnaeve G, et al (2020) End-to-end object detection with transformers. In: ECCV, pp 213–229

2020