Baseline Detection Model for Historical Persian Newspapers

Baseline Detection Model for Historical Persian Newspapers Summary childhood-in-the-iranian-press-persian-newspaper-baseline-v1.mlmodel detects text baselines in scanned, printed historical Persian newspapers. It is intended for use after page-region detection in a right-to-left OCR workflow. Its output is line geometry, not a transcription. Framework and compatibility Format: Kraken/Ketos .mlmodel Tested software environment: Kraken 5.3.0 Intended text direction: horizontal right-to-left Model SHA-256: 463edc7433ff918aea8ca003dc9493eece919c6914764b912df8515bc9cc2a13 Parent model and provenance The model was fine-tuned from Kraken's bundled default blla.mlmodel. The project parent file was byte-identical to the Kraken 5.3.0 packaged model: SHA-256 77a638a83c9e535620827a09e410ed36391e9e8e8126d5796a0f15b978186056. The upstream weight was introduced as Kraken's higher-resolution default segmentation model on 5 April 2022: https://github.com/mittagessen/kraken/blob/31f195ea5aa2276d70cc7bdc1de0a52fc166b51a/kraken/blla.mlmodel. The parent model is a general baseline/layout model, not documented as Persian-specific. No separate upstream model card, training-data description, DOI, or independently reported performance measure was located. Kraken is cited as Kiessling (2019); see THIRD-PARTY-NOTICES.md for the Apache-2.0 upstream notice. Training data persian-newspaper-baseline-alto-training-data-v1.zip contains 91 ALTO XML files from baseline-v1/, and no source images or non-ALTO files. The ALTO documents encode the training annotations but refer to source images that are not included. Training procedure The checked-in training procedure used Kraken Ketos baseline segmentation training, initialized from blla.mlmodel, with ALTO input, baseline-only training, image resizing, learning rate 0.0001, up to 1,000 epochs, early stopping, and lag 20. See TRAINING_COMMAND.txt. The model metadata records 81 completed training epochs. The exact completed command, random seed, GPU/hardware, split, and original training log are not documented for this release. Evaluation EVALUATION.md reports an in-sample class-independent baseline IoU of 0.721 over the 91 accompanying ALTO files (13,269 annotated baseline objects). This is not an estimate of generalization to unseen pages. Limitations and responsible use Results may degrade on damaged scans, highly irregular or multi-column layouts, tables, illustrations, decorative typography, and page designs absent from the annotation set. Inspect line placement against the page image before using OCR output as historical evidence. This model does not itself establish reading order or recognize text. Citation Sokol, L. 2026. Baseline Detection Model for Historical Persian Newspapers (Version 1.0.0) [machine-learning model]. Zenodo. https://doi.org/10.5281/zenodo.22537249