HPMD: A Historical Persian Manuscript Dataset for Word Spotting with Line-Level Annotation

Saeid Firouzi Daghigh; Majid Iranpour Mobarakeh

cs.CVSep 28, 2026

HPMD: A Historical Persian Manuscript Dataset for Word Spotting with Line-Level Annotation

Authors: Saeid Firouzi Daghigh, Majid Iranpour Mobarakeh

Organizations: Department of Computer Engineering and Information Technology, Payame Noor University, Tehran, Iran

Abstract

Large collections of historical Persian manuscripts have been digitized, but searching them is still slow and mostly manual. Historians usually want to find where a specific name, date, event, or topic appears, which is a word spotting problem. Progress on this task is limited by two things. First, there is almost no public dataset of historical Persian handwriting; the only notable resource, OpenITI MAKHZAN, contains a relatively small Persian portion. Second, word spotting models usually need word-level bounding boxes, which are very expensive to annotate. In this paper we introduce a new dataset of 223 pages, 3,678 lines, 37,631 words, and 130,630 characters, collected from diverse historical Persian books of poetry and prose and annotated at the region, line, and text level. We also propose a baseline that is trained only with line-level annotations but returns word-level locations. A fine-tuned line detector finds text lines, and a fine-tuned CRNN recognizer trained with CTC produces a frame-by-character posterior matrix for each line. Instead of decoding the most probable character at each frame, the query is scored directly against this matrix, so visually similar characters in Persian such as be and pe no longer cause hard failures. The frame alignment also gives the horizontal position of the word inside the line. On the test set, the fine-tuned line detector reaches an F1 of 0.892, and posterior-based search raises the word spotting F1 from 0.487 to 0.558 compared with exact matching on the decoded text, with the decision threshold selected on a held-out validation set. A PHOC attribute-embedding baseline that additionally receives oracle word boundaries at test time reaches an F1 of 0.449, below the proposed method. We also report a distributional analysis of the dataset, a taxonomy of retrieval errors, and a per-conditionbreakdown of performance.

Figures & tables

Dataset	Persian	Handwritten	Historical
VML-HD [ 6 ]	No	Yes	Yes
Urdu Katib [ 7 ]	No	Yes	Yes
IAM [ 5 ]	No	Yes	No
Sadri [ 1 ]	Yes	Yes	No
Khayyam [ 2 ]	Yes	Yes	No
Hoda [ 3 ]	Yes	Yes	No

TABLE I: Comparison of our dataset with related handwriting datasets.

Fig. 1: An example page from the dataset ( Majma al-Bayan ). Left: the original scanned page. Right: the annotation made in Transkribus, with text regions in blue, line regions in green, and the line transcriptions in red. The page contains a decorated frame, red rubrics, and a two-column poetry block, which illustrate the layout variety of the dataset.

Pages	Lines	Words	Characters
223	3,678	37,631	130,630

TABLE II: Overall statistics of the dataset.

Split	Pages	Lines	Words
Train	144	2,329	24,301
Validation	35	614	6,079
Test	44	735	7,251
Total	223	3,678	37,631

TABLE III: Page-level split of the dataset used for training and evaluation.

Quantity	Mean	Std	Median	Min	Max
Lines per page	16.4	6.3	15	1	30
Words per line	10.3	4.2	9	1	22
Characters per line	35.7	14.6	28	3	78
Characters per word	3.5	1.5	3	1	15

TABLE IV: Distribution of structural quantities in the dataset.

Fig. 2: Distribution of words per line (left), characters per word (middle), and lines per page (right). The bimodal shape of the first and third histograms comes from the mixture of two-column poetry pages and full-width prose pages.

Letter	Unicode	Count	Share (%)
zhe	U+0698	23	0.018
zal	U+0630	369	0.282
se	U+062B	394	0.302
za	U+0638	396	0.303
ghayn	U+063A	474	0.363
pe	U+067E	597	0.457

TABLE V: Frequency of the least frequent Persian letters after normalization. Counts are over 130,628 characters.

Fig. 3: Character frequency in the dataset after normalization, on a logarithmic scale. The distribution spans almost three orders of magnitude, from alef with about 17,000 occurrences to zhe with 23.

Attribute	Value	Pages	Lines	Words
Genre	poetry	133	2,535	22,032
	prose	84	1,033	14,660
	mixed	6	88	939
Style	shekasteh	173	3,000	32,878
	nastaliq	48	633	4,443
	naskh	2	23	310

TABLE VI: Composition of the dataset by genre, script style, degradation, illumination, and subjective reading difficulty.

Fig. 4: Composition of the dataset by genre, script style, degradation, illumination ( tazhib ), and subjective reading difficulty, in pages. The difficulty label coincides with the shekasteh script class, since that script drives the reading difficulty on almost every page where it appears.

Fig. 5: Overview of the proposed pipeline. Text lines are detected on the page, each line image is split into frames by the recognizer, and the recognizer outputs a frame-by-character posterior matrix (including the CTC blank). Word spotting is performed on this matrix instead of on the decoded text, and the matched frames give the position of the word in the line and the page.

Model	Precision	Recall	F1	Mean IoU
Base model	0.751	0.841	0.793	0.412
Our model	0.867	0.920	0.892	0.534

TABLE VII: Line detection results on the test set (44 pages, 735 lines) at an IoU threshold of 0.3.

Fig. 6: Effect of the decision threshold $\tau$ on the test set. Left: precision, recall, and F1 as functions of $\tau$ ; the dashed line marks $\tau^{*}=-0.80$ , selected by maximizing F1 on the validation set. Right: the corresponding precision-recall curve, with $\tau^{*}$ marked. Precision only rises sharply in the last fraction of the range, where recall has already collapsed.

Operating point	$\tau$	Precision	Recall	F1
Precision $\geq$ 0.90	$-0.02$	0.929	0.009	0.018
$\tau^{*}$ (max F1 on val)	$-0.80$	0.570	0.548	0.558
Recall $\geq$ 0.90	$-2.42$	0.071	0.900	0.132

TABLE VIII: Operating points of posterior search on the test set.

\tau^{*}

is selected by maximizing F1 on the validation set and is not tuned on test.

Model	Search	Precision	Recall	F1
Base model	Exact OCR	0.627	0.043	0.080
Base model	Posterior	0.288	0.218	0.248
Our model	Exact OCR	0.857	0.340	0.487
Our model	Posterior	0.570	0.548	0.558

TABLE IX: Word spotting results on the test set (44 pages, 735 lines, 7,251 words) with 396 query words. Posterior search uses

\tau^{*}=-0.80

selected on the validation set.

Level	Method	P	R	F1	mAP
Line	PHOC (oracle boxes)	0.664	0.339	0.449	0.266
Line	Posterior (ours)	0.570	0.548	0.558	0.564
Page	PHOC (oracle boxes)	0.601	0.375	0.462	0.393
Page	Posterior (ours)	0.601	0.639	0.619	0.639

TABLE X: Comparison with a PHOC attribute-embedding query-by-string baseline on the test set. The PHOC baseline additionally receives oracle word boundaries obtained by forced alignment of the ground-truth transcription, and is therefore an upper bound for word-box-supervised methods rather than a like-for-like comparison.

Query group	Queries	Precision	Recall	F1
Length 3-4	223	0.515	0.574	0.543
Length 5-6	144	0.861	0.524	0.651
Length 7+	17	1.000	0.444	0.615
Dotted $\leq$ 25%	118	0.580	0.577	0.578
Dotted 25-50%	175	0.555	0.567	0.561
Dotted $>$ 50%	91	0.540	0.535	0.538

TABLE XI: Word spotting performance by query property at

\tau^{*}=-0.80

Cause	Count	Share (%)
Substring of a longer word	339	53.4
Unrelated word	286	45.0
Same rasm , different dots	10	1.6

TABLE XII: Taxonomy of the 635 false positives of posterior search at

\tau^{*}=-0.80

Fig. 7: Qualitative examples of posterior search. (a), (b): false negatives. In (a) the query hadith is lost in a line with heavy ink blots; in (b) the short query qadah sits in a densely joined shekasteh hemistich where the recognizer spreads probability over several candidates. (c), (d): true positives with the word location recovered from the CTC frame alignment shown in red. No word-level annotation was used at any stage; the boxes come from the alignment alone.

Attribute	Value	Lines	P	R	F1
Tazhib	none	197	0.626	0.680	0.652
	border	446	0.522	0.446	0.481
	both	90	0.412	0.357	0.383
Difficulty	easy	59	0.518	0.496	0.507
	moderate	90	0.412	0.357	0.383
	hard	584	0.585	0.567	0.576

TABLE XIII: Word spotting performance on the test set (733 lines) grouped by page-level attribute, at

\tau^{*}=-0.80

. The difficulty and degradation groups are strongly imbalanced and confounded with script style; see the text.

Explore similar work

Apr 24, 2026cs.CV

Different Strokes for Different Folks: Writer Identification for Historical Arabic Manuscripts

Handwritten Arabic manuscripts preserve the Arab world's intellectual and cultural heritage, and writer identification supports provenance, authenticity verification, and historical analysis. Using the Muharaf dataset of historical Arabic manuscripts, we evaluate writer identification from individual line images and, to the best of our knowledge, provide the first baselines reported under both line-level and page-disjoint evaluation protocols. Since the dataset is only partially labeled for writer identification, we manually verified and expanded writer labels in the public portion from 6,858 (28.00%) to 21,249 lines (86.75%) out of 24,495 line images, correcting inconsistencies and removing non-handwritten text. After further filtering, we retained 18,987 lines (77.51%). We propose a Convolutional Neural Network (CNN)-based model with attention mechanisms for closed-set writer identification, including rare two-writer lines modeled as composite writer-pair classes. We benchmark fourteen configurations and conduct ablations across different feature extractors and training regimes. To assess generalization to unseen pages, the page-disjoint protocol assigns all lines from each page to a single split. Under the line-level protocol, a fine-tuned DenseNet201 with attention achieves 99.05% Top-1 accuracy, 99.73% Top-5 accuracy, and 97.44% F1-score. Under the more challenging page-disjoint protocol, the best observed results are 78.61% Top-1 accuracy, 87.79% Top-5 accuracy, and 66.55% F1-score, thus quantifying the impact of page-level cues. By expanding the Muharaf dataset's labeled subset and reporting both protocols, we provide a clearer benchmark and a practical resource for historians and linguists engaged with culturally and historically significant documents. The code and implementation details are available on GitHub.

Hamza A. Abushahla, Ariel Justine N. Panopio, Layth Al-Khairulla +1

Department of Computer Science and Engineering, American University of Sharjah, Sharjah 26666, United Arab Emirates

Handwritten Text Recognition Authorship Arabic Document Segmentation Textual Dataset Open-Set Recognition+2

Ask PDF

Jul 22, 2026cs.CV

Persian Pixel: A large-scale synthetic OCR dataset for Persian language

Optical Character Recognition (OCR) for Persian remains substantially less mature than for Latin-script languages despite Persian being spoken by more than 110 million people across multiple countries. This gap arises from two fundamental challenges: the intrinsic complexity of the Perso-Arabic writing system and the limited availability of large-scale, high-quality annotated datasets. Persian script exhibits obligatory cursive connectivity, context-dependent glyph shaping, extensive ligatures, diacritic placement, and stylistic variation across writing forms such as Naskh and Nastaliq, all of which significantly complicate text recognition. At the same time, the high cost and labor-intensive nature of manual annotation have created a persistent data bottleneck, limiting the development of robust OCR systems and slowing progress in Persian document digitization.In this paper, we introduce Persian Pixel, a comprehensive synthetic OCR dataset specifically designed to address these challenges. Comprising over 343,000 high-fidelity image text pairs, the dataset spans sentence, paragraph, and full-page document layouts generated from a carefully curated seven-million-word Persian corpus using the SynthOCR-Gen rendering framework. The generation pipeline faithfully models the typographic characteristics of Persian script, including contextual character joining, positional glyph variants, diacritic placement, and multiple representative Persian typefaces. To bridge the synthetic-to-real domain gap, the rendered images are further enriched with more than twenty-five stochastic degradation models that emulate realistic document acquisition artifacts, including ink bleed, paper aging, blur, illumination variation, scanner imperfections, compression artifacts, and multiple noise processes.By overcoming the long-standing scarcity of annotated Persian OCR data, Persian Pixel provides a scalable and openly available resource for training and fine-tuning modern OCR architectures, including transformer-based models such as TrOCR and Donut. The dataset establishes a strong foundation for research in Persian document analysis, historical manuscript digitization, and end-to-end document understanding, while demonstrating that programmatic synthetic data generation offers a practical, cost-effective, and scalable alternative to manual annotation for advancing OCR in low-resource and typographically complex scripts.

Pouria Mahdi, Haq Nawaz Malik

Persian Optical Character Recognition Handwritten Text Recognition Few-Shot Font Generation Synthetic-To-Real Domain Gap

Ask PDF

Sep 14, 2026cs.CL

ParsHate: A Benchmark Dataset for Hate and Target Detection in Persian

We introduce ParsHate, a manually annotated dataset of 10,000 Persian tweets spanning 2013-2022, representing the first decade-long benchmark for hate speech detection in Persian. The dataset contains 31% hateful content and supports both hate detection and multi-label fine-grained target identification across seven structured target categories. ParsHate also distinguishes explicit and implicit hate, marks explicit and implicit targets, and provides span-level rationales. Data collection combines random and score-stratified temporal sampling to reduce keyword-driven bias while preserving natural label distributions. Applying SOTA models for Persian hate-speech detection on ParsHate shows moderate performance (79% F1), especially with samples from earlier years, and low performance with target identification (25.5% macro-F1). This emphasizes the diverse sampling of hate speech in ParsHate and its challenging nature that requires more advanced methods for better performance. Dataset is made publicly available.

Zahra Bokaei, Walid Magdy, Bonnie Webber

School of Informatics University of Edinburgh

Hate Speech Detection Hate Speech Persian Natural Language Datasets Benchmark Datasets

Ask PDF