PaperlayerView exact revision on arXiv

Source-linked AI summary

Frugal Handwritten Text Recognition for Sub-Saharan Civil-Status Registers: A Smartphone-to-Records Pipeline with CPU-Only Inference

Ousmane Hamadou, Jean Louis Kedieng Ebongue Fendji

arXiv:2610.36540v1cs.CV

TL;DR

Historical handwritten registers are often photographed but remain unsearchable, while local administrations face domain and resource constraints. We assess a consumer-capture pipeline with CPU inference and a fine-tuned recognizer; it sharply lowers CER in one registry, though writer overlap limits generalization.

  • Problem

    Historical handwritten registers remain unsearchable by name, date, or place, and available HTR models mismatch local handwriting, templates, and names while often assuming resources local offices lack.

    Source: Introduction · p. 2

  • Method

    The pipeline combines consumer smartphone capture, classical CPU preprocessing, and a Kraken recognizer fine-tuned on annotated records, with CPU-only inference.

    Source: Introduction · p. 2, §4.2 · p. 8

  • Results

    Fine-tuning reduced CER from 14.53% to 0.50% on corrected segmentation and seen writers; the recognizer runs under 6 s per image and under 2.6 GiB on a laptop CPU.

    Source: Introduction · p. 2, §8 · p. 13

  • Takeaways & Limitations

    Laptop-CPU deployment may avoid GPUs or cloud services at inference time, although adapting the model to a new registry currently requires GPU access.

    Source: §6 · p. 12, §8 · p. 13

    The evaluation comes from one registry, with possible writer overlap across training and test and a test set of about 28 records, limiting evidence about unseen writers or sites.

    Source: §6 · p. 12

Abstract

from arXiv · show

Handwritten civil-status registers across Sub-Saharan Africa are being photographed, but rarely transcribed, which leaves their content unsearchable. Local administrations that hold these records typically lack GPUs, reliable connectivity, and the budget for external digitization services. We study whether a frugal pipeline, with consumer-grade capture and CPU-only inference, can make such registers machine-readable. We introduce a corpus of 1,568 smartphone images of handwritten birth, marriage, and death records from a municipal registry in Chad, with line-level transcriptions for 185 records. Our pipeline combines a fully classical preprocessing stage, running at about 2.6 images per second on a laptop CPU, with a Kraken recognizer fine-tuned from a pretrained model in about four hours on a free-tier GPU. On manually corrected line segmentation and a single-registry test set whose writers also appear in training, fine-tuning reduces the character error rate from 14.53% to 0.50% and the word error rate from 56.93% to 3.37%. On a six-core laptop CPU, the recognizer weighs 61 MiB, peaks below 2.6 GiB of RAM, and processes an image in under 6 s on average, excluding line segmentation. Finally, a downstream information-extraction task shows that residual recognition errors cost about 9 F1 points on role resolution, compared with a 36.7-point gap between the fine-tuned and pretrained recognizers. These results are evidence of feasibility within one registry, not of performance on unseen writers or sites; we discuss these limitations in detail.

Introduction

Historical handwritten civil-status registers remain difficult to search, while local administrations face domain mismatch and resource constraints. This paper evaluates a consumer-capture, CPU-inference pipeline and reports recognition results under explicitly limited conditions.

Introduction

  • Historical registers remain unsearchable by name, date, or place, while available HTR models mismatch local handwriting, templates, and names and often require unavailable compute or connectivity.These records can be the only surviving proof of identity or filiation for people without recent documents.Source: Introduction · p. 2
  • The corpus contains 1,568 smartphone images of birth, marriage, and death records from a Chad municipal registry, with line-level transcriptions for 185 records.The records use five administrative templates.Source: Introduction · p. 2
  • A classical preprocessing pipeline handles page detection, record separation, margin removal, and upside-down detection at about 2.6 images per second on a laptop CPU.It ran on the corpus without errors, but output accuracy was not measured.Source: Introduction · p. 2
  • 14.53% to 0.50% CER after fine-tuning, on manually corrected line segmentation and a test set whose writers also appear in training.The model was fine-tuned for about four hours on a free-tier GPU; inference runs on a laptop CPU.Source: Introduction · p. 2
  • The paper analyzes how recognition errors propagate into a downstream information-extraction task.The analysis estimates how much end-task error comes from recognition rather than later stages.Source: Introduction · p. 2
  • Capture, preprocessing, and inference use consumer hardware without a GPU, but line segmentation was corrected manually, so the pipeline was not evaluated end to end.The study also identifies a single collection site, a small test set, and no inter-annotator agreement as limitations.Source: Introduction · p. 3

2 Related Work

Prior work spans segmentation-free HTR, historical-record recognition and extraction, and efficient deployment. The paper addresses the relative lack of comparable resources for Sub-Saharan African civil registers and targets local-administration constraints.

2 Related Work

  • Segmentation-free line recognition uses CTC to learn image-to-character alignment; Kraken implements a lightweight CRNN-CTC design suited to CPU inference.The passage contrasts this with higher-cost encoder-decoder Transformers such as TrOCR.Source: §2 · p. 3
  • Historical-record work includes marriage licences, an extraction competition, and French civil registers, but comparable Sub-Saharan African civil-register resources are absent.The paper highlights varied handwriting, mixed printed and handwritten forms, and local names as distinctions from European corpora.Source: §2 · p. 3
  • Pretrained language models may underrepresent Sub-Saharan African names, while the recognition-model concern is plausible but not measured on name spans.The paper measures adaptation effects by entity type in extraction, not on recognizer name spans.Source: §2 · pp. 3–4
  • Civil-registration programmes mainly computerize new registrations or index scanned records, while this work targets historical handwritten content with a recipe sized for local administrations.The deployment framing follows Green AI's call to report quality alongside computational cost.Source: §2 · p. 4

3 Dataset and Capture Protocol

The dataset comes from a municipal civil-status archive in Chad, where handwritten records preserve past registrations and exhibit substantial variation in writers, templates, and physical condition. The collection and annotation protocol uses consumer equipment and records important coverage and access constraints.

3 Dataset and Capture Protocol

  • The corpus comes from a municipal registry in Chad outside the urban areas targeted by recent reforms; handwritten registers remain the only record of many past events.The source describes the registry as representative of local offices where paper registers are the sole copy.Source: §3 · p. 4
  • Successive officers wrote without a common calligraphic standard; records mix printed forms, sometimes bilingual French–Arabic, with French handwriting and local names.Physical wear includes stains, tears, faded ink, and humidity or handling damage.Source: §3 · p. 4
  • The five-day collection produced 1,568 images using a smartphone, stand, and laptop for framing and remote triggering.The phone captured RAW images at 2268×4032 pixels.Source: §3 · p. 4
  • RAW capture avoids phone processing that can destroy intensity variations used by the cropping pipeline; parallel positioning and remote triggering limit blur and perspective distortion.Lighting was uncontrolled, and the full setup cost about 150,000 FCFA (roughly 230 EUR).Source: §3 · p. 5
  • Birth records dominate the image corpus and marriage records are scarce; five templates include three structurally different birth layouts.The variation requires handling differing line lengths, field arrangements, and printed–handwritten mixes.Source: §3 · p. 5, Table 1 · p. 5
  • Line-level transcriptions cover 185 records, with a 70/15/15 line split constrained so lines from one record stay together; annotation was done by one annotator.Purely printed lines are excluded, while printed text sharing a line with handwriting is included; inter-annotator agreement is unavailable.Source: §3 · p. 5
  • The corpus is not publicly released because it contains personal information; preprocessing code is planned for release, while recognizer weights await checks and authorization.Source: §3 · p. 5

4 Method

The method combines classical preprocessing, fine-tuned handwritten text recognition, and a downstream extraction probe, with inference on a laptop CPU. The stages are evaluated separately, including recognition on manually corrected line boundaries.

4 Method

  • The three-stage pipeline preprocesses smartphone captures, recognizes handwriting with fine-tuned Kraken, and uses downstream extraction only to probe recognition-error propagation; inference runs on a laptop CPU.The extraction stages are not part of the evaluated vision pipeline.Source: §4 · pp. 5–6, Figure 1 · p. 6

4.1 Classical preprocessing

  • 2.6 images per second: classical preprocessing ran across all 1,568 images without errors, but crop, split, and orientation accuracy was not measured.Its hand-tuned parameters may need retuning elsewhere.Source: §4.1 · pp. 6–7

4.2 Handwritten text recognition

  • Recognition uses a hybrid convolutional-recurrent Kraken model; fine-tuning merges the pretrained and corpus alphabets and takes about four hours on a free-tier GPU.CER and WER are computed on manually corrected, ground-truth line boundaries, excluding segmentation errors.Source: §4.2 · pp. 7–8

4.3 Cost measurement protocol

  • The cost protocol compares pretrained and fine-tuned models at FP32 and measures model size, loading and warm-up, memory, timing, throughput, CPU use, and training cost.Recognition timing uses stored corrected line segmentation; reduced-precision inference was not evaluated.Source: §4.3 · p. 8

4.4 Downstream extraction (evaluation probe)

  • The extraction probe assigns roles using deterministic, template-specific rules, so role errors can arise from recognition, NER, or template classification.The probe fine-tunes a French NER model on annotated records before role assignment.Source: §4.4 · pp. 8–9

5 Experiments and Results

On manually corrected line segmentation, fine-tuning sharply improves recognition on a test set with writer overlap, while CPU benchmarks show modest resource demands. Downstream results also show that residual recognition errors affect role resolution, alongside errors in the extraction stage.

5.1 Setup

  • CER counts character substitutions, deletions, and insertions relative to reference characters; WER is defined analogously at word level.The experimental setup defines these metrics for reporting recognition quality.Source: §5.1 · p. 9

5.2 Recognition quality

  • 0.50% CER versus 14.53% for the pretrained model, with fine-tuning reducing WER by about 17×, on manually corrected segmentation and writers also seen in training.The authors treat this as evidence of feasibility within one registry, not an estimate for unseen writers or other registries.Source: §5.2 · p. 9

5.3 Frugality of the recognizer

  • About 10 images per minute and 3.5–3.7 lines per second were processed on a six-core laptop CPU; both models stayed below 2.6 GiB peak memory.The benchmark covered 50 images in five steps and 1,037 lines; the authors do not interpret small cost differences between the models.Source: §5.3 · p. 10, Table 4 · p. 10
  • Fine-tuning took about four hours on a free-tier GPU, the only pipeline step requiring a GPU; the 0.50% CER met the authors’ assumed 2% ceiling.The 2% threshold is a working criterion, subject to the recognition-quality caveats.Source: §5.3 · pp. 10–11

5.4 How recognition errors propagate

  • Even with perfect input text, role resolution reaches only 90.3% F1; residual recognition errors cost about 9 F1 points, while the pretrained recognizer loses a further 36.7 points.The extraction stage was fixed and tested on 46 records outside the HTR test set; the fine-tuned recognizer may have seen some records during training.Source: §5.4 · p. 11, Table 5 · p. 11

5.5 Downstream extraction (summary)

  • The fine-tuned NER model reached 98.0% micro-F1 over eight entity types on ground-truth transcriptions; template classification reached 87.0% accuracy, with G2/G3 confusion not harming role resolution.NER was evaluated on a held-out split of 1,242 entities; template classification was evaluated on 100 records.Source: §5.5 · p. 11

6 Discussion and Limitations

The observed CER gain is limited by confounded factors and writer overlap, while several deployment costs and evaluation boundaries constrain the study’s claims. The authors identify testing across writers and sites and measuring additional cost dimensions as important next steps.

6 Discussion and Limitations

  • The 29-fold CER reduction does not explain why adaptation helps: baseline type, character coverage, and overlap with training writers are confounded.The improvement on local names is qualitative; the proposed explanation that generic models poorly suit local names remains a hypothesis to test.Source: §6 · pp. 11–12
  • About four hours of GPU training and unmeasured human annotation effort are required to adapt the model; CPU training was not evaluated.The resulting model can run on a laptop CPU, but lower-end machines and reduced-precision inference were not tested.Source: §6 · p. 12
  • The 0.50% CER may overestimate performance on unseen writers or other registries because records come from one site and officers or registers can overlap across splits.The test set has about 28 records; segmentation was manually corrected, timings exclude segmentation, and no confidence intervals or repeated runs are reported.Source: §6 · p. 12

7 Ethics Statement

The study uses real, identifying civil-status records, often without the subjects’ knowledge, and does not anonymize or pseudonymize the working corpus. The authors describe limits on data handling and publication, but note both a cloud-upload exception and unresolved regulatory questions.

7 Ethics Statement

  • The corpus contains identifying information, and most people whose records were used were not informed of the research.The working corpus was not anonymized or pseudonymized; training and evaluation used original identifying text.Source: §7 · p. 12
  • Data were processed locally except for fine-tuning, which uploaded training images and transcriptions to a free-tier cloud GPU; the data and annotations are not released.The paper shows only two isolated text lines with names and record identifiers masked, not full records.Source: §7 · p. 13
  • The authors do not claim that the research was declared to or authorized by ANSICE, or that public-archive research is exempt from individual consent.The Municipality of Koumra gave written permission for academic use, required masking of published personal information, and excluded commercial use.Source: §7 · p. 13
  • Local CPU-only deployment under the record-holding administration’s control reduces surveillance or profiling risk compared with outsourced processing, but does not remove it.The paper identifies searchability of identity records at scale as a possible support for surveillance or profiling.Source: §7 · p. 13

8 Conclusion

The paper tests a frugal pipeline for making handwritten civil-status registers machine-readable and reports a large recognition improvement after fine-tuning. Its feasibility evidence comes from one registry, with writer overlap and a small test set, and the authors outline further evaluation and privacy work.

8 Conclusion

  • Fine-tuning reduced CER from 14.53% to 0.50% on corrected line segmentation and writers seen in training, using about four hours of GPU training.On a laptop CPU, the recognizer runs in under 6 s per image and uses under 2.6 GiB of memory.Source: §8 · p. 13
  • The downstream extraction task showed that residual recognition errors cost about 9 F1 points on role resolution.Source: §8 · p. 13
  • The findings suggest deployment-time recognition need not require GPUs or cloud services, but do not establish performance beyond one registry or on unseen writers.The test set is small and uses a single baseline.Source: §8 · p. 13
  • Future work includes evaluation across registers, writers, and sites, end-to-end testing with automatic line segmentation, and an anonymization scheme before sharing data or models.The authors also propose measuring handwritten-text and name error rates, annotation-size effects, annotator agreement, and comparisons with additional baselines.Source: §8 · p. 13

Related papers

Loading 2610.36540v1…