Learning based Ge'ez character handwritten recognition

Federico Cunico; Hailegabriel Dereje Degefa; Hailemicael Lulseged Yimer; Marco Cristani

REVIEW 4 major objections 4 minor 17 references

Learning based Ge'ez character handwritten recognition

T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A CNN–LSTM recognizer trained on the HHD-Ethiopic dataset claims state-of-the-art handwritten Ge'ez recognition, with OOD CER 26.95 and NED 26.50, below all eight compared systems and below human performance.

desk verdict New numbers on a public benchmark that beat prior Ge'ez OCR baselines, but the abstract's two-stage method doesn't match the described CRNN, so the SOTA claim isn't supported as written. read the letter →

arxiv 2411.13350 v1 pith:6XM4VCO3 submitted 2024-11-20 cs.CV

Hailemicael Lulseged Yimer , Hailegabriel Dereje Degefa , Marco Cristani , Federico Cunico This is my paper · ORCID

classification cs.CV

keywords Ge'ezscripthandwrittentextrecognitionCNN-LSTMconnectionisttemporalclassificationEthiopicOCRHHD-Ethiopicdatasetcharactererrorratelow-resource

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 argues that a hybrid CNN–LSTM recognizer can set a new best result for handwritten Ge'ez word recognition on the HHD-Ethiopic dataset, beating eight published OCR systems and the reported human-level baseline. The proposed pipeline is dual-stage: a CNN is trained to recognize individual Ge'ez characters and then acts as a feature extractor for an LSTM word-level recognizer, trained with Connectionist Temporal Classification so no character boundary labels are needed. On the in-distribution test set the paper reports CER 15.48 and NED 15.00, and on the harder out-of-distribution 18th-century manuscript test set it reports CER 26.95 and NED 26.50. The authors present this as a step toward digitizing historical Ethiopic manuscripts and as evidence that low-resource ancient scripts can be served by a relatively standard CNN–LSTM+CTC design.

What carries the argument

The load-bearing mechanism is a CNN–LSTM sequence recognizer trained with Connectionist Temporal Classification on the 182-character Ethiopic syllabary. Four residual convolutional blocks, with channels growing from 64 to 512, extract hierarchical visual features and reduce spatial dimensions through max pooling, with dropout and batch normalization for regularization; the resulting feature map feeds two stacked bidirectional LSTM layers of 512 units per direction, and a fully connected layer outputs character probabilities per time step plus a blank CTC token. The abstract's 'dual-stage' framing adds a separate CNN character-recognition stage whose weights then serve as the feature extractor for the LSTM word recognizer. The evaluation machinery is the HHD-Ethiopic dataset's IID split (6,375 samples) and OOD split (15,935 historical samples), scored by Character Error Rate and Normalized Edit Distance.

What would settle it

Reproduce the experiment under one protocol: download the released model and code, retrain each of the eight baselines on the same HHD-Ethiopic IID/OOD splits, and compute CER/NED on the OOD set. Any baseline scoring below 26.95 CER or 26.50 NED, or a rerun of the proposed model producing a higher value than 26.95, would disprove the state-of-the-art claim.

Watch

Extended reading notes

Core claim

The paper claims that its CNN–LSTM architecture reaches state-of-the-art performance for handwritten Ge'ez recognition on the HHD-Ethiopic dataset, surpassing eight compared methods and the human-level baseline. The decisive reported numbers are CER 26.95 and NED 26.50 on the out-of-distribution test set, against 28.65 and 27.37 for the best prior baseline and 33.20 and 33.70 for human readers. The design intent is that the convolutional blocks extract visual features from characters and the stacked bidirectional LSTMs model the sequential dependencies of words, with CTC allowing learning from unsegmented image–text pairs. The authors present the result as an advance for digitizing historical Ethiopic manuscripts and for low-resource script recognition.

Load-bearing premise

The paper's superiority claim stands on the assumption that every baseline in Table I was trained and evaluated on the same HHD-Ethiopic splits with the same preprocessing and metric calculation as the proposed model; if the baselines came from different protocols, the CER/NED gaps would not be comparable.

Editorial extensions

If this is right

Handwritten Ge'ez can be transcribed without character-level segmentation, since CTC training aligns unsegmented image sequences to word or line transcripts.
The benchmark gap between IID (CER 15.48) and OOD (CER 26.95) gives a quantified measure of how far historical-manuscript style diverges from modern handwriting.
If the results reproduce, the HHD-Ethiopic human-performance baseline is no longer the upper bound for this task, and future systems should be compared against this CNN–LSTM result.
The architecture is not script-specific beyond the output character set, so the same dual-stage recipe can be lifted directly to other Ethiopic-script languages such as Amharic.

Reading between the lines

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

The paper does not ablate the two stages, so I would not infer from it alone that the separate character-level CNN pretraining is what drives the gain; an end-to-end CTC model without that stage could perform equally well.
The reported margin over the strongest prior baseline (OOD CER 28.65 to 26.95) is about 1.7 points; whether that translates into practically useful transcription depends on the error distribution across common and rare characters, which the paper does not report.
Because the baselines' training details are not described, the fair inference is that the architecture works well on this dataset; the comparative ranking's strength will be settled only by reproducing all methods in one protocol.
A testable extension the authors do not run: apply the same dual-stage pipeline to other low-resource historical scripts with syllabic writing systems and small corpora, using their own CNN-first-then-LSTM schedule.

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

The useful thing here is a new set of numbers on the HHD-Ethiopic benchmark: CER 15.48 IID, 26.95 OOD, NED 26.50 OOD, which beat the HPopt-Attn-CTC baseline from Belay et al. If those numbers hold up, that is a real step for a low-resource script. The authors also do something right: they evaluate on the same public splits, report IID/OOD separately, and include the human-performance row from [8]. That is the right way to frame a niche OCR result.

The soft spot is not minor. The abstract and contributions claim a two-stage pipeline: first train a CNN on isolated characters, then use it as a frozen feature extractor for an LSTM word recognizer. The Methods section, III-A and III-B, describes a single end-to-end CRNN with four residual CNN blocks, two BiLSTMs, and CTC loss. No character-level pretraining step is described. There is one sentence in Section IV about the 'character recognition model' loss/accuracy and a figure, but no architecture or training details. So the reader cannot tell what was actually evaluated. The copy-paste slip in III-A ('recognize Amharic text') adds to the impression that the write-up was rushed.

Also missing: code, training epochs, learning rate, optimizer, preprocessing, and any error bars or repeated runs. Table I lists baseline numbers but the caption just says 'trained on HHD [8]' with no protocol. The numbers may be perfectly comparable, but the paper gives no reason to trust that they were run under identical conditions. That is a comparability concern, though I don't see evidence of fabricated numbers: this is a plausibly correct CRNN on a public dataset.

So the empirical result is new and could matter to the Ethiopic OCR community, but the central claim as stated is unsupported by the method. The paper should not be accepted in this form. It deserves a serious referee, not a desk reject, because the benchmark question is real and the issues are fixable in revision. If the two-stage claim is false and it's just a CRNN, the authors need to say so and reposition the contribution as a strong CRNN baseline. If the two-stage system exists, they need to specify it fully. Either way, the letter should be a revise-with-major-changes, not a clean reject.

Referee Report

4 major / 4 minor

Summary. The manuscript proposes a CNN-LSTM system for handwritten Ge'ez word recognition, claiming a two-stage process in which a CNN is first trained to recognize individual characters and then used as a feature extractor for an LSTM-based word recognizer. The authors report Character Error Rate (CER) and Normalized Edit Distance (NED) on the HHD-Ethiopic dataset under both in-distribution (IID) and out-of-distribution (OOD) test sets, and claim state-of-the-art results, including outperforming human performance. Section IV presents comparisons with SVTR, ASTER, ABINet, CRNN, Plain-CTC, Attn-CTC, and two HPopt variants.

Significance. If the results are verified, a strong handwritten-text recognition system for the low-resource Ge'ez script would be a meaningful contribution, with clear applications to historical document digitization and cultural-heritage preservation. The paper uses a publicly available benchmark, standard evaluation metrics, and includes human-performance comparison. However, the current manuscript does not provide enough evidence to verify the central claim: the architecture described in the Methods section is a conventional single-stage CRNN, no code or training details are supplied, and the baseline numbers appear to be imported from prior work without a documented evaluation protocol. These issues are load-bearing because the stated novelty and the state-of-the-art claim are both defined by the two-stage design.

major comments (4)

[Abstract and Section I, versus Section III] The abstract and the contribution bullet in Section I claim a 'two-stage recognition process': first a CNN is trained to recognize individual characters, then it acts as a feature extractor for an LSTM-based word recognizer. However, Section III.A and Section III.B describe only a single end-to-end CRNN: four residual CNN blocks, two stacked bidirectional LSTM layers, and CTC training. No step of separately training a character classifier is described, and the only trace of such a model is the passing mention in Section IV of the 'loss and classification accuracy of the character recognition model' associated with Figure 4. The reader therefore cannot determine what system produced the numbers in Table I, and the claimed state-of-the-art results cannot be attributed to the stated contribution. The authors should either provide a complete description of the two-stage training procedure or revise the abstract and contribution list to describe the actual CRNN system.
[Section IV, Table I] The baseline comparison in Table I is not sufficiently specified. The caption says only 'trained on HHD [8]', and the row values for Human-performance, SVTR, ASTER, ABINet, CRNN, Plain-CTC, Attn-CTC, HPopt-Plain-CTC, and HPopt-Attn-CTC appear to be taken directly from reference [8] or other cited works. The manuscript does not describe whether the same preprocessing, same IID/OOD splits, same training data, same decoding method, or same random seeds were used for all methods. Without this information, the reported margins (for example, OOD CER 26.95 versus 28.65 for the best prior method) may simply reflect differences in evaluation protocol rather than algorithmic improvements. The authors should re-run the baselines under a shared protocol or provide a detailed description of how each baseline number was obtained, together with error bars or confidence intervals.
[Section IV and Table I] There is a direct numerical inconsistency in the reported results. The text in Section IV states that 'our model obtains 15.48 for CER using the IID test set, and 27.96 for the OOD test set', but Table I lists the OOD CER for 'Ours' as 26.95, and the abstract and contribution bullet also report a CER of 26.95 and an NED of 26.50. The authors should reconcile these values and ensure that all reported numbers correspond to the same test split and metric definition. In addition, Eq. (2) defines NED as a value between 0 and 1, but the values in Table I are reported as percentages (e.g., 26.50), so the scaling convention should be clarified.
[Section III.B and IV] The experimental setup is not reproducible from the information given. Section III.B mentions only that training and evaluation were performed on a Linux system with an NVIDIA RTX 3090 GPU; it does not report the optimizer, learning rate, batch size, number of epochs, input image size, data augmentation, or random seeds. No code or trained model is provided in the submission. Because the central claim is a new state of the art on a public benchmark, the authors should provide the code and/or checkpoint, or at a minimum a complete training and evaluation protocol, including how many independent runs were performed and how variability was measured.

minor comments (4)

[Section III.A] The section says the model is designed to 'recognize Amharic text in images', while the title and abstract concern Ge'ez. Although Amharic and Ge'ez share the Ethiopic script, the terminology should be consistent and clearly distinguish the target language/script.
[Section III.B] There is a typo: 'A Finally, a fully connected layer' should read 'Finally, a fully connected layer'. Also, the phrase 'maps the attended features' refers to attention, but no attention mechanism is described anywhere in the architecture.
[Section IV, Figure 4] Figure 4 is said to show 'the loss and classification accuracy of the character recognition model', but the caption and text do not explain what this character recognition model is, how it was trained, or how it relates to the word-level CRNN described in Section III. The figure also lacks axis labels and a clear caption.
[Section IV] The sentence describing the HHD-Ethiopic dataset contains the fragment 'with in the testing distribution', which appears to be incomplete. It also says the dataset comprises 79,684 samples including two test sets; it would be clearer to state the number of training samples separately from the IID and OOD test sets.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed SOTA rests on an external benchmark and standard metrics, with no load-bearing reduction to fitted inputs or self-citations.

full rationale

The paper's central claim is an empirical CER/NED comparison on the HHD-Ethiopic dataset. The reported numbers are computed with the standard CER and NED definitions (Eqs. 1 and 2) from predicted and ground-truth sequences, and the comparison targets (SVTR, ASTER, ABINet, CRNN, Plain-CTC, Attn-CTC, HPopt variants, and human performance) are taken from external published work, principally [8], not from the authors' own prior results. There is therefore no fitted parameter that is later renamed a prediction, and no equation in which the claimed output is defined in terms of the input. The only self-citations ([16], [17]) appear in the conclusion as future research directions and are not load-bearing for the recognition result. The discrepancy between the abstract's 'two-stage' description and the single-stage CRNN described in Section III is a reproducibility/specification concern, not a circularity: nothing in the derivation chain shows the result being assumed by construction. Accordingly, the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper's central claim depends on the reliability of the external dataset and baselines, plus a set of unspecified training details. No new entities are introduced.

free parameters (5)

CNN channels progression (64 to 512)
Chosen by hand; not reported as tuned, but the architecture's capacity affects the reported performance.
Bi-LSTM hidden units (512)
Chosen by hand; not justified or ablated.
Dropout rate
Mentioned as applied, but no rate given.
Number of training epochs
Figure 4 shows loss curves but no epoch count or stopping criterion is stated.
Learning rate and optimizer
Not specified in the implementation details.

assumptions (4)

domain assumption HHD-Ethiopic dataset provides ground-truth labels and the IID/OOD splits are exactly as described in [8].
The evaluation relies entirely on this dataset and its splits, which the paper does not verify.
domain assumption Baseline models in Table I were trained/evaluated under comparable conditions to the proposed model.
Needed for the SOTA claim; not demonstrated in the paper.
domain assumption CTC loss is an appropriate objective for the unsegmented text-line images.
A standard assumption in OCR, but not tested here.
domain assumption The architecture described in Section III is the one that produced the numbers in Table I.
The abstract's two-stage description contradicts the method, so this is an assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning based Ge'ez character handwritten recognition." pith.science (2026). https://pith.science/paper/6XM4VCO3

@misc{pith2026241113350,
  author       = {Pith},
  title        = {Pith review of: Learning based Ge'ez character handwritten recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6XM4VCO3}},
  note         = {Machine review of arXiv:2411.13350}
}

read the original abstract

Ge'ez, an ancient Ethiopic script of cultural and historical significance, has been largely neglected in handwriting recognition research, hindering the digitization of valuable manuscripts. Our study addresses this gap by developing a state-of-the-art Ge'ez handwriting recognition system using Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks. Our approach uses a two-stage recognition process. First, a CNN is trained to recognize individual characters, which then acts as a feature extractor for an LSTM-based system for word recognition. Our dual-stage recognition approach achieves new top scores in Ge'ez handwriting recognition, outperforming eight state-of-the-art methods, which are SVTR, ASTER, and others as well as human performance, as measured in the HHD-Ethiopic dataset work. This research significantly advances the preservation and accessibility of Ge'ez cultural heritage, with implications for historical document digitization, educational tools, and cultural preservation. The code will be released upon acceptance.

Figures

Figures reproduced from arXiv: 2411.13350 by the authors.

**Figure 2.** Challenges in Ge’ez handwriting recognition. (a) and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Examples of predictions made by the model. (a) [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: The loss and classification accuracy of the character [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 11 canonical work pages

[8]

Hhd-ethiopic a historical handwritten dataset for ethiopic ocr with baseline models and human-level performance,

B. H. Belay, I. Guyon, T. Mengiste, B. Tilahun, M. Liwicki, T. Tegegne, and R. Egele, “Hhd-ethiopic a historical handwritten dataset for ethiopic ocr with baseline models and human-level performance,” 2023

work page 2023
[1]

Online and off-line handwriting recognition: a comprehensive survey,

R. Plamondon and S. N. Srihari, “Online and off-line handwriting recognition: a comprehensive survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 22, no. 1, pp. 63–84, 2000

work page 2000
[2]

Gradient-based learning applied to document recognition,

Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, 1998

1998
[3]

A study on the performance of convolutional neural networks in offline handwritten amharic character recognition,

A. Moges, A. Bewketu, D. Dadi, and T. Dibaba, “A study on the performance of convolutional neural networks in offline handwritten amharic character recognition,” in 2017 International Conference on Advances in Computing, Communications and Informatics (ICACCI) , pp. 1117–1123, IEEE, 2017

work page 2017
[4]

Imagenet classification with deep convolutional neural networks,

A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Communications of the ACM , vol. 60, no. 6, pp. 84–90, 2017

2017
[5]

Long short-term memory,

S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

1997
[6]

Data augmentation for recognition of handwritten words and lines using a cnn-lstm network,

C. Wigington, S. Stewart, B. Davis, B. Barrett, B. Price, and S. Cohen, “Data augmentation for recognition of handwritten words and lines using a cnn-lstm network,” in 2017 14th IAPR international conference on document analysis and recognition (ICDAR) , vol. 1, pp. 639–645, IEEE, 2017

work page 2017
[7]

Ge’ez language,

Encyclopedia Britannica, “Ge’ez language,” 2023. Accessed on August 21, 2024

work page 2023

Show all 17 references

[9]

Amharic ocr: an end-to-end learning,

B. Belay, T. Habtegebrial, M. Meshesha, M. Liwicki, G. Belay, and D. Stricker, “Amharic ocr: an end-to-end learning,” Applied Sciences , vol. 10, no. 3, p. 1117, 2020

2020
[10]

Ocr for amharic scripts using deep learning,

B. Belay et al. , “Ocr for amharic scripts using deep learning,” African Journal of Information Systems , vol. 12, no. 3, pp. 223–244, 2020

2020
[11]

Svtr: Scene text recognition with a single visual model,

Y . Du, Z. Chen, C. Jia, X. Yin, T. Zheng, C. Li, Y . Du, and Y .-G. Jiang, “Svtr: Scene text recognition with a single visual model,” arXiv preprint arXiv:2205.00159, 2022

2022 arXiv
[12]

An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition,

B. Shi, X. Bai, and C. Yao, “An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition,” IEEE transactions on pattern analysis and machine intelligence, vol. 39, no. 11, pp. 2298–2304, 2016

2016
[13]

Aster: An attentional scene text recognizer with flexible rectification,

B. Shi, M. Yang, X. Wang, P. Lyu, C. Yao, and X. Bai, “Aster: An attentional scene text recognizer with flexible rectification,” IEEE transactions on pattern analysis and machine intelligence , vol. 41, no. 9, pp. 2035–2048, 2018

2018
[14]

Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition,

S. Fang, H. Xie, Y . Wang, Z. Mao, and Y . Zhang, “Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 7098–7107, 2021

2021
[15]

A blended attention-ctc network architecture for amharic text-image recognition.,

B. H. Belay, T. Habtegebrial, M. Liwicki, G. Belay, and D. Stricker, “A blended attention-ctc network architecture for amharic text-image recognition.,” in ICPRAM, pp. 435–441, 2021

2021
[16]

A machine learning-oriented survey on tiny machine learning,

L. Capogrosso, F. Cunico, D. S. Cheng, F. Fummi, and M. Cristani, “A machine learning-oriented survey on tiny machine learning,” IEEE Access, 2024

2024
[17]

I-split: Deep network interpretability for split computing,

F. Cunico, L. Capogrosso, F. Setti, D. Carra, F. Fummi, and M. Cristani, “I-split: Deep network interpretability for split computing,” in 2022 26th International Conference on Pattern Recognition (ICPR) , pp. 2575– 2581, IEEE, 2022

2022

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [8]

Hhd-ethiopic a historical handwritten dataset for ethiopic ocr with baseline models and human-level performance,

B. H. Belay, I. Guyon, T. Mengiste, B. Tilahun, M. Liwicki, T. Tegegne, and R. Egele, “Hhd-ethiopic a historical handwritten dataset for ethiopic ocr with baseline models and human-level performance,” 2023

work page 2023

[2] [1]

Online and off-line handwriting recognition: a comprehensive survey,

R. Plamondon and S. N. Srihari, “Online and off-line handwriting recognition: a comprehensive survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 22, no. 1, pp. 63–84, 2000

work page 2000

[3] [2]

Gradient-based learning applied to document recognition,

Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, 1998

1998

[4] [3]

A study on the performance of convolutional neural networks in offline handwritten amharic character recognition,

A. Moges, A. Bewketu, D. Dadi, and T. Dibaba, “A study on the performance of convolutional neural networks in offline handwritten amharic character recognition,” in 2017 International Conference on Advances in Computing, Communications and Informatics (ICACCI) , pp. 1117–1123, IEEE, 2017

work page 2017

[5] [4]

Imagenet classification with deep convolutional neural networks,

A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Communications of the ACM , vol. 60, no. 6, pp. 84–90, 2017

2017

[6] [5]

Long short-term memory,

S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

1997

[7] [6]

Data augmentation for recognition of handwritten words and lines using a cnn-lstm network,

C. Wigington, S. Stewart, B. Davis, B. Barrett, B. Price, and S. Cohen, “Data augmentation for recognition of handwritten words and lines using a cnn-lstm network,” in 2017 14th IAPR international conference on document analysis and recognition (ICDAR) , vol. 1, pp. 639–645, IEEE, 2017

work page 2017

[8] [7]

Ge’ez language,

Encyclopedia Britannica, “Ge’ez language,” 2023. Accessed on August 21, 2024

work page 2023