Crossing Language Borders: A Pipeline for Indonesian Manhwa Translation

Nithyasri Narasimhan; Sagarika Singh

REVIEW 5 major objections 6 minor 12 references

Crossing Language Borders: A Pipeline for Indonesian Manhwa Translation

T0 review · 5 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A three-stage pipeline that chains fine-tuned YOLOv5xu bubble detection, Tesseract OCR, and MarianMT translation can translate Indonesian Manhwa panels into English automatically.

desk verdict A clear course-project pipeline for Indonesian Manhwa translation with plausible component scores, but the end-to-end translation claim is not actually measured and the comparison section overreaches. read the letter →

arxiv 2501.01629 v1 pith:STJ6XBNK submitted 2025-01-03 cs.LG cs.CLcs.CV

Nithyasri Narasimhan , Sagarika Singh This is my paper · ORCID

classification cs.LGcs.CLcs.CV

keywords IndonesianManhwatranslationspeechbubbledetectionYOLOv5TesseractOCRMarianMTlow-resourcelanguagetextoverlaymachine

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

The paper claims that a fully automatic pipeline can translate Indonesian Manhwa (Korean comics translated into Indonesian) into English by chaining three fine-tuned components: YOLOv5xu to find speech bubbles, Tesseract OCR with the Indonesian language model to read the text, and MarianMT to translate it, with the translated text laid back over the panels. The authors report component-level results: F1 90.7% for bubble detection, 3.1% character error rate and 8.6% word error rate for OCR, and BLEU 0.27 with METEOR 0.61 for translation. If these numbers hold, the work shows that a practical, low-cost baseline for low-resource comic translation can be assembled from existing models and a relatively small dataset (538 annotated images for detection, plus existing parallel corpora for translation). The significance is in the workflow: tasks that typically take days by hand are reduced to hours, and the approach is presented as transferable to other underrepresented language pairs.

What carries the argument

The pipeline itself is the central mechanism: (1) fine-tuned YOLOv5xu detects speech bubbles and gives bounding boxes; (2) Tesseract OCR with the Indonesian language model (OEM 3, PSM 6) reads the extracted bubbles, with grayscale preprocessing; (3) fine-tuned MarianMT, trained on Identic and OpenSubtitles, translates the text; (4) OpenCV and Pillow overlay the translated text back into the bubble shapes. The load-bearing identity is the coupling of these components; each step's output becomes the next step's input, so the reported component scores are only as strong as the assumption that the test panels are representative.

What would settle it

Take a held-out set of Indonesian Manhwa panels spanning multiple series, artists, and image qualities (including stylized fonts, overlapping text, and low resolution), run the pipeline on them, and compute OCR CER/WER against human transcription and translation BLEU/METEOR against professional references. If CER and WER rise substantially compared to the reported 3.1% and 8.6% on ordinary panels, or if the end-to-end translation loses speech bubbles, the claim of strong component performance is not representative.

Watch

Extended reading notes

Core claim

The central discovery is that a domain-specific, three-stage pipeline automatically translates Indonesian Manhwa panels to English while preserving context and artistic layout, despite the low-resource setting. The paper shows that fine-tuning a pre-trained YOLOv5xu on a 538-image webcomics dataset yields reliable speech-bubble detection (F1 90.7%), that applying Tesseract's Indonesian model directly to extracted bubbles gives low character error (3.1% CER, 8.6% WER) when panels are clean, and that fine-tuning MarianMT on Identic plus OpenSubtitles produces translations that retain meaning (METEOR 0.61) better than n-gram fidelity (BLEU 0.27) would suggest. The complete pipeline, including text overlay with OpenCV and Pillow, produced translated panels that the authors judge to keep context and meaning, demonstrating feasibility of automating Manhwa translation for a low-resource language pair.

Load-bearing premise

The reported OCR error rates assume the test panels are clean and the stylized font difficulty is low, because the paper does not describe how OCR ground truth was selected, which panels were tested, or how many samples produced the 3.1% CER and 8.6% WER.

Editorial extensions

If this is right

A practical baseline exists for automating low-resource comic translation: each component works well enough to support the end-to-end workflow.
Translators and scanlation teams can cut per-chapter time from days to hours by using such a pipeline, though human editing would still be needed.
A relatively small detection dataset (538 images) is sufficient to fine-tune a detector for speech bubbles in Manhwa, suggesting similar results are achievable for other comic styles.
Combining a formal parallel corpus (Identic) with a conversational one (OpenSubtitles) is a viable recipe for fine-tuning a translation model for dialogue-heavy content.
The same architecture could be adapted to other low-resource language pairs by swapping the OCR language pack and the parallel corpus.

Reading between the lines

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

The paper does not measure end-to-end quality on an entire chapter; a fair evaluation would track how errors propagate from missed bubbles to OCR mistakes to translation drift, since a missed bubble removes dialogue entirely.
The OCR ground-truth procedure is undocumented, so the 3.1% CER may not generalize to the stylized fonts and noisy backgrounds common in Manhwa; a controlled font/background study would define the boundary of the claim.
The component comparisons in Section V are indirect, pitting each component against different baselines on different benchmarks; a direct head-to-head on the same test set would be needed to assert superiority.
The pipeline's structure (detection → OCR → MT → overlay) is language-agnostic; the same recipe could be tried for other Southeast Asian languages, but the MarianMT step would need a similarly matched conversational parallel corpus.

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 read: this is a course-project pipeline paper, and it is exactly as modest and as limited as it looks. The one genuinely new thing is the specific combination — YOLOv5xu fine-tuned on webcomic bubble detection, Tesseract with Indonesian, MarianMT fine-tuned on OpenSubtitles+Identic — applied to Indonesian-to-English Manhwa. I don't know of a prior pipeline for that language pair, and the authors don't cite one. The system description is clear, the component numbers are plausible, and the paper is honest about the absence of a dedicated Manhwa dataset and about informal-language difficulties.

The soft spots are concentrated in the evaluation. The stress-test note is right: Section IV-C's BLEU 0.27 / METEOR 0.61 are never tied to OCR output from Manhwa panels. They look like they were computed on the clean parallel test split from OpenSubtitles/Identic. That means the end-to-end claim — that the pipeline translates comics acceptably — rests on visual inspection of a few panels, not on a measured pipeline evaluation. For a course project that is forgivable; as a published claim it is a real gap. The OCR numbers have the same problem: no description of ground-truth creation, test panel selection, or sample size, and CER 3.1% is surprisingly clean for stylized fonts. Section V's comparisons are not apples-to-apples: CO-DETR on COCO, YOLOX-L on Manga109-s, and the OCR baseline on different comic data are different tasks and domains; saying 'superior performance' overstates what the numbers show. The authors do mention dataset differences, but the tone of the section is more confident than the evidence.

What is good: the work is clearly explained, the pipeline is a sensible integration, and the limitations are acknowledged up front. It could serve as a starting point for someone wanting a baseline for low-resource comic translation. It is not reproducible as submitted — no code, no data, no error bars — but it is not incoherent or deceptive.

My take: this is a fine student report that should not be treated as a rigorous research result. If it comes to a venue, I would not desk-reject it out of hand for a workshop or student-track venue; the integration is useful and the evaluation gaps are fixable with a modest revision that reports OCR ground truth, links MT scores to actual OCR output, and drops or re-frames the non-comparable comparisons. For a main conference, I'd send it to reviewers only if the authors release artifacts and address those gaps. As is, it is not referee-worthy at a strong venue.

Referee Report

5 major / 6 minor

Summary. The paper proposes an automated Indonesian-to-English Manhwa translation pipeline combining fine-tuned YOLOv5xu for speech bubble detection, Tesseract OCR with the Indonesian language model, fine-tuned MarianMT for machine translation, and an overlay stage for reintegrating translated text into panels. The authors report component-level scores on a small Webcomics dataset and on OpenSubtitles/Identic parallel data: detection F1 90.7%, mAP@0.5:0.95 88.9%, OCR CER 3.1% and WER 8.6%, and MT BLEU 0.27 / METEOR 0.61. They conclude that the pipeline successfully automates Manhwa translation, with end-to-end success supported by visual inspection of sample outputs. The paper also compares its components against prior work and claims consistently superior performance.

Significance. If the reported component numbers were firmly established and linked to a measured end-to-end evaluation, the paper would provide a useful practical baseline for low-resource comic translation. The authors choose a relevant language pair and use standard, publicly available datasets, with held-out splits that avoid circular evaluation. However, the current evidence does not support the central claim of a working end-to-end pipeline: the MT scores are not tied to OCR-extracted Manhwa text, the OCR evaluation protocol is under-specified, the comparative baselines are not matched on datasets, and the only end-to-end evidence is visual inspection. The contribution is therefore a plausible component-level proof of concept rather than a validated system, and the paper would need substantial additional evaluation to justify its main conclusion.

major comments (5)

[§IV-C and §II] The BLEU 0.27 and METEOR 0.61 scores are not shown to be computed on OCR-extracted Manhwa speech-bubble text. The dataset section describes only an 80/10/10 split of combined OpenSubtitles and Identic data for training/validation/testing, and no statement connects that test set to the pipeline's actual input. Since OCR errors and the informal, stylized dialogue of Manhwa differ from clean general-domain parallel text, these scores do not establish translation quality on the pipeline's real input. Please report MT metrics on the OCR output of a held-out set of Manhwa panels, or clearly justify why the clean-text test set is representative, and state the test-set size for every reported metric.
[§III-B and §IV-B] The OCR evaluation lacks a specified protocol: the paper does not state how many bubbles or panels were tested, how the ground truth was created, which chapters or sources were sampled, or whether the evaluated text was actually stylized Manhwa text. Given the pipeline target, the reported CER 3.1% and WER 8.6% are strikingly low compared to published comic OCR results (e.g., the segmentation-free method in [11] reports CER 22.78% and WER 39.30%), and the claim that images were 'of satisfactory quality' is not quantified. Please provide the ground-truth collection procedure, the number of test samples, and error bars or a per-sample distribution, so that the reader can assess representativeness.
[§V] The comparative claims in Section V are not supported because the baselines are evaluated on different datasets and tasks: CO-DETR on COCO, YOLOX-L on Manga109-s, Rigaud et al. on comic books (not necessarily Manhwa), and Dwiastuti on IWSLT spoken-language data. A comparison across different test sets cannot establish 'consistently superior performance' as claimed. Please either run the baseline methods on the same evaluation data used for the proposed pipeline or explicitly reframe the numbers as contextual references rather than as a head-to-head comparison.
[§II and §IV-A] The dataset description is internally inconsistent: the paper states that the Roboflow Webcomics dataset contains 538 images, but then reports a split of 465 training and 118 validation images, which sums to 583. Additionally, no test split is described for the detection model, so the F1 score, mean precision, recall, and mAP values in Table I lack a clearly defined evaluation set. Please reconcile the dataset size and specify the exact training/validation/test split used when generating Table I.
[§IV-D] The end-to-end claim rests solely on 'visual inspection of translated samples' with no quantitative or structured human evaluation. Since the central contribution is the complete pipeline, the accumulation of detection, OCR, and translation errors should be measured on a held-out set of panels, for example through task-specific metrics (e.g., final-panel text accuracy) or a structured human rating protocol. Without such evidence, the reported component scores cannot substantiate the statement that the pipeline 'successfully integrates all steps.'

minor comments (6)

[§I] There is a typo in Section I: 'detecting and extracting speech bubbles bubbles' should read 'detecting and extracting speech bubbles.'
[§III-B and references] The Tesseract citation is inconsistent: the abstract and introduction cite Tesseract as [2], but Section III-B cites reference [7], which is actually the automatic manga text detection paper by Zhang et al. Please unify the citation for Tesseract.
[Table I] The notation 'Mean mAP@0.5' is confusing; the text later refers to mAP@0.5:0.95, and the relationship between mAP@0.5 = 96.3%, mean mAP@0.5 = 88.9%, and F1 = 90.7% should be clarified. Also, no confidence intervals or error bars are reported for any metric in Tables I–III.
[§IV-D] The text references 'Fig. 7' as an example of the final translated image, but the manuscript contains only Figures 2 and 3. Please add the figure or correct the cross-reference.
[§V-C and Table III] The BLEU score is reported as 0.27 in Table III but as 27% in Section V-C; please use a single consistent scale when comparing with the 23% baseline.
[§II and §III-C] The paper does not describe the fine-tuning procedure for MarianMT (e.g., hyperparameters, number of epochs, base model checkpoint) or state whether code and trained models will be released. Providing these details is important for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline claims are engineering evaluations, not derivations, and no reported score reduces to its own input by construction.

full rationale

The paper makes no theoretical derivation and contains no prediction that is defined in terms of its own inputs. Each component is evaluated with standard external metrics: YOLOv5xu is fine-tuned on a Roboflow Manhwa dataset and scored with precision, recall, mAP, and F1; Tesseract OCR is scored with CER and WER; MarianMT is trained on the Identic and OpenSubtitles corpora with an 80/10/10 split and scored with BLEU and METEOR. None of these scores are fitted constants renamed as predictions, and the paper does not invoke any self-citation or uniqueness theorem to force its choices. The acknowledged limitations in Section VI, such as the lack of a dedicated Manhwa dataset and challenges with informal language, are exactly the kind of external-data representativeness concerns that affect accuracy claims, not circularity. Likewise, the unspecified OCR ground truth and the unclear connection between the MT test set and actual speech-bubble text are validity risks, but they do not make the reported numbers equivalent to the paper's inputs by construction. Cross-dataset comparisons in Section V are methodologically questionable but not circular. Therefore the appropriate finding is no significant circularity.

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

The central claim depends on assumptions about dataset representativeness and evaluation validity, not on explicit hand-fitted constants or invented theoretical entities. The model weights are learned from data, but no ad hoc free parameters are introduced beyond the standard fine-tuning choices.

assumptions (4)

domain assumption YOLOv5xu can be fine-tuned effectively on a 538-image dataset to detect Manhwa speech bubbles.
Section III-A assumes the Roboflow Webcomics Text Selection dataset annotations are correct and general enough for Manhwa panels, despite the small size and unverified annotation quality.
domain assumption Tesseract with the Indonesian language model, OEM 3, PSM 6, and grayscale preprocessing is adequate for stylized Manhwa fonts.
Section III-B asserts strong OCR performance on stylized fonts and complex layouts but provides no separate evaluation of font difficulty or image quality.
domain assumption BLEU and METEOR scores computed on an undisclosed test subset measure translation quality.
Section III-E relies on these metrics, but no test set size, reference source, or confidence intervals are given.
domain assumption Indonesian is a low-resource language for which MarianMT fine-tuned on Identic and OpenSubtitles is a suitable translation approach.
Section II and Section III-C assume these corpora, combined without documented proportions, are representative of Manhwa dialogue.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Crossing Language Borders: A Pipeline for Indonesian Manhwa Translation." pith.science (2026). https://pith.science/paper/STJ6XBNK

@misc{pith2026250101629,
  author       = {Pith},
  title        = {Pith review of: Crossing Language Borders: A Pipeline for Indonesian Manhwa Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/STJ6XBNK}},
  note         = {Machine review of arXiv:2501.01629}
}

read the original abstract

In this project, we develop a practical and efficient solution for automating the Manhwa translation from Indonesian to English. Our approach combines computer vision, text recognition, and natural language processing techniques to streamline the traditionally manual process of Manhwa(Korean comics) translation. The pipeline includes fine-tuned YOLOv5xu for speech bubble detection, Tesseract for OCR and fine-tuned MarianMT for machine translation. By automating these steps, we aim to make Manhwa more accessible to a global audience while saving time and effort compared to manual translation methods. While most Manhwa translation efforts focus on Japanese-to-English, we focus on Indonesian-to-English translation to address the challenges of working with low-resource languages. Our model shows good results at each step and was able to translate from Indonesian to English efficiently.

Figures

Figures reproduced from arXiv: 2501.01629 by the authors.

**Figure 1.** Overview of Methodology A. Speech Bubble Detection We used pre-trained Yolov5xu [1], and fine-tuned it further on Manhwa dataset [4]. YOLOv5xu is a lightweight, efficient object detection model that excels in identifying objects in images with high precision and recall. We used this fine-tuned model to detect bounding boxes of speech bubbles and extract the dimensions of the bounding boxes and extract these speech b… view at source ↗

**Figure 2.** Side-by-side comparison of the original panel and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Pipeline steps: (a) Enhanced speech bubbles, (b) Final [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages

[11]

Segmentation-free speech text recognition for comic books,

C. Rigaud, J.-C. Burie, and J.-M. Ogier, “Segmentation-free speech text recognition for comic books,” in 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR) , vol. 03, 2017, pp. 29–34. 4

work page 2017
[1]

Ultralytics yolov5,

G. Jocher, “Ultralytics yolov5,” 2020. [Online]. Available: https: //github.com/ultralytics/yolov5

2020
[2]

Tesseract OCR Engine,

R. Smith, “Tesseract OCR Engine,” 2007. [Online]. Available: https://web.archive.org/web/20160819190257/tesseract-ocr. googlecode.com/files/TesseractOSCON.pdf

work page arXiv 2007
[3]

Marian: Fast neural machine translation in c++,

M. Junczys-Dowmunt, R. Grundkiewicz, T. Dwojak, H. Hoang, K. Heafield, T. Neckermann, F. Seide, U. Germann, A. Fikri Aji, N. Bogoychev et al., “Marian: Fast neural machine translation in c++,” in Proceedings of ACL: System Demonstrations , 2018, pp. 116–121

work page 2018
[4]

Webcomics text selection dataset,

L. B. Nunes, “Webcomics text selection dataset,” https://universe. roboflow.com/luciano-bastos-nunes/webcomics-text-selection, dec 2024, visited on 2024-12-10. [Online]. Available: https: //universe.roboflow.com/luciano-bastos-nunes/webcomics-text-selection

work page 2024
[5]

Identic: The indonesian-english parallel corpus,

R. M. Rosa, I. Warsono, P. E. Murni, W. Ariyanti, S. K. Ika, and M. Adri- ani, “Identic: The indonesian-english parallel corpus,” in Proceedings of the 8th Workshop on Asian Language Resources (ALR 2010) , 2010, pp. 71–74

work page 2010
[6]

Opensubtitles2016: Extracting large parallel corpora from movie and tv subtitles,

P. Lison and J. Tiedemann, “Opensubtitles2016: Extracting large parallel corpora from movie and tv subtitles,” in Proceedings of the 10th International Conference on Language Resources and Evaluation (LREC 2016), 2016, pp. 923–929

work page 2016
[7]

Towards Fully Automated Manga Translation

G. Zhang, X. Yan, M. Iwamura, and Y . Matsuo, “Automatic manga text detection and recognition,” arXiv preprint arXiv:2012.14271v2 , 2021

work page Pith review arXiv 2012

Show all 12 references

[8]

Efficient and accurate scene text detector,

Y . Baek, B. Lee, D. Han, S. Yun, and H. Lee, “Efficient and accurate scene text detector,” arXiv preprint arXiv:2103.14027v3 , 2021

2021 arXiv
[9]

Detrs with collaborative hybrid assign- ments training,

Z. Zong, G. Song, and Y . Liu, “Detrs with collaborative hybrid assign- ments training,” 10 2023, pp. 6725–6735

2023
[10]

Usb: Universal-scale object detection benchmark,

Y . Shinya, “Usb: Universal-scale object detection benchmark,” 03 2021

2021
[12]

English-Indonesian neural machine translation for spoken language domains,

M. Dwiastuti, “English-Indonesian neural machine translation for spoken language domains,” in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, F. Alva-Manchego, E. Choi, and D. Khashabi, Eds. Florence, Italy: A...

2019

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [11]

Segmentation-free speech text recognition for comic books,

C. Rigaud, J.-C. Burie, and J.-M. Ogier, “Segmentation-free speech text recognition for comic books,” in 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR) , vol. 03, 2017, pp. 29–34. 4

work page 2017

[2] [1]

Ultralytics yolov5,

G. Jocher, “Ultralytics yolov5,” 2020. [Online]. Available: https: //github.com/ultralytics/yolov5

2020

[3] [2]

Tesseract OCR Engine,

R. Smith, “Tesseract OCR Engine,” 2007. [Online]. Available: https://web.archive.org/web/20160819190257/tesseract-ocr. googlecode.com/files/TesseractOSCON.pdf

work page arXiv 2007

[4] [3]

Marian: Fast neural machine translation in c++,

M. Junczys-Dowmunt, R. Grundkiewicz, T. Dwojak, H. Hoang, K. Heafield, T. Neckermann, F. Seide, U. Germann, A. Fikri Aji, N. Bogoychev et al., “Marian: Fast neural machine translation in c++,” in Proceedings of ACL: System Demonstrations , 2018, pp. 116–121

work page 2018

[5] [4]

Webcomics text selection dataset,

L. B. Nunes, “Webcomics text selection dataset,” https://universe. roboflow.com/luciano-bastos-nunes/webcomics-text-selection, dec 2024, visited on 2024-12-10. [Online]. Available: https: //universe.roboflow.com/luciano-bastos-nunes/webcomics-text-selection

work page 2024

[6] [5]

Identic: The indonesian-english parallel corpus,

R. M. Rosa, I. Warsono, P. E. Murni, W. Ariyanti, S. K. Ika, and M. Adri- ani, “Identic: The indonesian-english parallel corpus,” in Proceedings of the 8th Workshop on Asian Language Resources (ALR 2010) , 2010, pp. 71–74

work page 2010

[7] [6]

Opensubtitles2016: Extracting large parallel corpora from movie and tv subtitles,

P. Lison and J. Tiedemann, “Opensubtitles2016: Extracting large parallel corpora from movie and tv subtitles,” in Proceedings of the 10th International Conference on Language Resources and Evaluation (LREC 2016), 2016, pp. 923–929

work page 2016

[8] [7]

Towards Fully Automated Manga Translation

G. Zhang, X. Yan, M. Iwamura, and Y . Matsuo, “Automatic manga text detection and recognition,” arXiv preprint arXiv:2012.14271v2 , 2021

work page Pith review arXiv 2012