PaperlayerView exact revision on arXiv

Source-linked AI summary

MuSTEL: A Large-Scale Multilingual Synthetic Text-Line Image Dataset for OCR

Majid Adibian, Hossein Saberi, Hossein Zeinali

arXiv:2610.34148v1cs.CV

TL;DR

Multilingual line-level OCR must accommodate differing writing directions, typography, and visual conditions, but training-data diversity remains important and resources jointly covering Persian, Arabic, and English are limited. MuSTEL constructs a large synthetic line-image dataset using multilingual text, capability-aware fonts, mixed-direction rendering, and visual variation. Its broad language and mixed-language coverage supports research on multilingual line recognition, while its synthetic, line-level scope does not reproduce all real-document properties or page layouts.

  • Problem

    Line-level OCR requires training examples spanning languages, writing directions, typography, and visual appearance, while resources jointly covering Persian, Arabic, and English remain limited.

    Source: §1 · p. 2

  • Method

    MuSTEL generates annotated text-line images from multilingual corpora using validated capability-aware fonts, bidirectional rendering, within-line font and style changes, and varied visual conditions.

    Source: §3.4 · pp. 10–11, §6 · p. 24

  • Results

    MuSTEL provides broad coverage across Persian, Arabic, and English, including mixed-language samples, and is suitable for research on multilingual and mixed-direction line recognition in document-oriented OCR pipelines.

    Source: §6 · p. 24

  • Takeaways & Limitations

    The dataset provides a resource for training and evaluating multilingual line-level OCR systems and investigating recognition across languages, writing directions, and visual conditions.

    Source: §6 · p. 24

    MuSTEL is synthetic and line-level, so it does not fully reproduce real scanned or photographed documents or model complete page layouts and other document-structure problems.

    Source: §5.3 · p. 23

Abstract

from arXiv · show

Optical character recognition (OCR) systems depend strongly on the scale and diversity of their training data. One approach to document OCR first analyzes the page layout and extracts text-line images, which are then transcribed by a recognition model. Multilingual line-level recognition involving Persian, Arabic, and English presents additional challenges because different writing directions, fonts, and styles may coexist within a single line. This paper introduces MuSTEL, a large-scale synthetic multilingual OCR dataset containing more than 14.5 million annotated text-line images, organized into a training partition and predefined small and large test sets. Samples are generated from natural-language corpora, including multilingual content already present in the source texts. The generation framework incorporates capability-aware font selection, within-line font and style changes, mixed-direction text rendering, highlighting, and underlining. A validated repository of 2,009 font files from 711 families supports broad typographic coverage, while varied backgrounds, blur, geometric distortion, skew, stroke width, and image save quality provide diverse document-like appearances. Statistical analysis characterizes the dataset's language composition, transcription lengths, lexical and character-sequence diversity, image dimensions, typography, and rendering conditions. MuSTEL provides a resource for training and evaluating multilingual line-level OCR systems and investigating recognition robustness across languages, writing directions, and visual conditions.

1. Introduction

OCR effectiveness depends on the scale, quality, and diversity of training data, while line-level recognition requires examples that reflect language, writing direction, typography, and image variation. MuSTEL addresses the limited availability of Persian-Arabic-English line-level resources with a large synthetic dataset and varied rendering.

1. Introduction

  • Recognition effectiveness and generalization depend strongly on training-data scale, quality, and diversity, and synthetic generation can produce accurately labeled line images at scale.Multilingual document OCR also calls for varied scripts, fonts, styles, and visual conditions.Source: §1 · p. 2
  • Line-level OCR separates transcription from page-layout analysis, but requires training data spanning language, writing direction, typography, and image appearance.This setting supports pipelines that detect or segment lines before recognizing them.Source: §1 · p. 2
  • Persian and Arabic run right-to-left while English runs left-to-right, so mixed-language lines require recognition to handle direction changes while preserving transcription order.Lines may also vary in font, style, size, emphasis, numerals, punctuation, backgrounds, blur, skew, distortion, and compression.Source: §1 · p. 2
  • Existing line-level resources jointly covering Persian, Arabic, and English remain limited, as do datasets combining multilingual content with broad typographic and visual variation.MuSTEL provides more than 14.5 million annotated line images, mixed-language content, predefined partitions, and varied fonts, styles, and document-like appearances.Source: §1 · pp. 2–3

2. Related Work

OCR datasets vary in image source, annotation granularity, language coverage, and rendering. MuSTEL’s distinction is its combination of Persian, Arabic, and English line images with mixed-language composition and controlled within-line font and style changes.

Dataset types and recognition granularity

  • Scene-text and document-oriented datasets differ by image source, while character-, word-, line-, region-, and page-level annotations support different tasks and are not directly interchangeable.Page-level resources require line segmentation and transcriptions before training a line recognizer; line-level pairs support evaluation of sequences with punctuation, spaces, and script or font transitions.Source: §2 · p. 3

Existing OCR datasets

  • MJSynth and SynthText primarily target English scene-text or word recognition, leaving limited representation of multilingual document lines, right-to-left scripts, and within-line changes.Arabic and Persian resources differ in recognition unit, scale, and typographic coverage; most remain language-specific.Source: §2 · p. 4

Multilingual datasets

  • Multilingual benchmarks may include Arabic and English without Persian or without transitions between right-to-left and left-to-right scripts within a recognition sequence.MuSTEL targets Persian-Arabic-English lines with direction changes and controlled font-family and style variation.Source: §2 · p. 4

Synthetic data generation

  • TRDG offers configurable rendering, but its standard workflow lacks dedicated within-line language and font-style switching, highlighting, and underlining.MuSTEL extends the framework to support mixed-language lines with font-family and style changes, underlining, and highlighting.Source: §2 · p. 5

Dataset comparison

  • None of the datasets listed in Table 1 jointly covers Persian, Arabic, and English; MuSTEL combines that coverage with more than 14.5 million line images and varied document-like rendering.Its distinction also includes mixed-language composition and controlled within-line changes in language, font family, and style.Source: Table 1 · p. 5, §2 · p. 6

3. Dataset Construction

MuSTEL extends TRDG into a synthetic line-image pipeline combining multilingual corpora, validated fonts, script-aware rendering, and document-like visual variation. Its construction also includes automated annotations and quality checks, with predefined test sets at two evaluation scales.

3.1. Overview of the Dataset Generation Pipeline

  • The generation pipeline samples source text, selects compatible fonts and styles, renders text, varies image appearance, and checks image, transcription, and metadata before acceptance.Mixed-language lines come from multilingual content already present in the source corpora, and text is preserved without normalization.Source: §3.1 · pp. 6–7

3.2. Font Collection and Validation

  • The font repository uses automated and manual checks to validate script-specific character rendering, including joining, ligatures, hamzah forms, and diacritics.Capability labels are recorded independently, allowing partially supported fonts to be selected only for compatible text segments.Source: §3.2 · pp. 8–9

3.3. Text Corpus Collection and Preprocessing

  • Persian, Arabic, and English corpora combine Wikipedia articles and digitized books, with markup and duplicate lines removed where applicable.Long texts are divided into randomly sized segments for line generation; segments may contain part of a sentence rather than a complete one.Source: §3.3 · p. 9

3.4. Synthetic Line-Image Generation

  • MuSTEL extends TRDG with language-aware font assignment, bidirectional composition, script-specific shaping, and controlled within-line font-family and style changes.Mixed-direction lines preserve logical transcription order, while compatible fonts are checked against each run’s required characters before rendering.Source: §3.4 · p. 10

3.5. Diversity in Data Generation and Quality Control

  • Rendering varies font size, text appearance, margins, backgrounds, blur, skew, geometric distortion, stroke width, image quality, highlighting, and underlining.Samples receive their own combinations of text, fonts, backgrounds, geometric effects, and image-quality settings, recorded in metadata.Source: §3.5 · pp. 11–12

3.6. Dataset Organization and Splits

  • The dataset provides a 6,000-image small test set and a 143,999-image large test set, alongside a training set of 14,377,989 images.The small set has 2,000 samples from each primary language; the large set is approximately balanced across English, Persian, and Arabic.Source: §3.6 · p. 13

4. Dataset Statistics and Analysis

The analysis characterizes MuSTEL's scale and multilingual balance alongside textual, typographic, and image-appearance diversity. It reports measurements of language composition, transcription and image dimensions, lexical and character-sequence coverage, and rendering variation.

4.1. Overall Dataset Statistics

  • MuSTEL contains 14,527,988 images, with English, Persian, and Arabic each comprising about one-third; its training, small-test, and large-test partitions contain 14,377,989, 6,000, and 143,999 images.The small test set has 2,000 samples per language; the large test set has 48,000 English, 47,999 Persian, and 48,000 Arabic samples.Source: §4.1 · p. 14, Table 2 · p. 14

4.3. Image Statistics

  • The validated repository contains 2,009 font files from 711 families, every file and family appears in the dataset, and Arial accounts for less than 3% of assignments.Usage is distributed across a broad range of families; Figure 10 shows the most frequently used families.Source: §4.1 · p. 14, §4.3 · pp. 18–19, Figure 10 · p. 20

4.2. Textual Statistics

  • Monolingual samples make up most of the dataset, while 1,871,326 samples (12.88%) contain more than one language in a line.Language composition uses source-language labels and character-level transcription analysis because Persian and Arabic share many Unicode characters.Source: §4.2 · pp. 14–15, Table 4 · p. 15
  • Mean transcription lengths are 34.67 Persian, 35.76 Arabic, 41.77 English, and 45.08 mixed-language characters, with an overall mean of 38.71.Lengths use Unicode grapheme clusters, counting letters with attached marks as one unit and including whitespace.Source: §4.2 · pp. 15–16
  • Across the complete dataset, MuSTEL contains more than 103 million word occurrences and approximately 5.2 million unique words; its character n-gram analysis reaches 675,457 unique trigrams.Source: §4.2 · pp. 16–17

4.3. Image Statistics

  • Image widths have a 354-pixel median and a 198–592-pixel central range, while heights have a 36-pixel median and a 31–42-pixel central range.Transcription length and width are positively related (Pearson correlation 0.652), though width also depends on script and rendering choices.Source: §4.3 · p. 18, Figure 8 · p. 18, Table 7 · p. 18
  • Most samples use one font family and style, while approximately 8% include two families or two styles; Regular and Bold account for most style assignments.The style analysis consolidates recorded labels into seven broader categories.Source: §4.3 · pp. 19–20, Table 8 · p. 20, Table 9 · p. 21
  • Plain white backgrounds are most common and quasicrystal backgrounds least common; reduced quality and blur are prevalent, geometric distortion affects about half, and highlighting and underlining each appear in roughly 7%.Rendering-condition percentages are not mutually exclusive because multiple variations can occur in one image.Source: §4.3 · pp. 21–22, Figure 11 · p. 21, Table 10 · p. 22

5. Discussion

MuSTEL combines multilingual line-level scale with typographic and visual diversity, supporting OCR training, evaluation, and robustness research. Its synthetic, line-only scope leaves real-document conditions and page-level tasks incompletely represented.

5.1. Advantages of MuSTEL

  • MuSTEL provides 14,527,988 text-line images with broadly balanced Persian, Arabic, and English representation, including 1,871,326 mixed-language samples.Its combination of three-language coverage, mixed-language lines, typographic diversity, and rendering variation distinguishes the dataset's stated strengths.Source: §5.1 · p. 22, §5 · p. 22
  • All 2,009 validated font files from 711 families appear in the dataset, with capability-aware selection and within-line font-family and style changes.The repository covers validated support for letters, numerals, punctuation, hamzah forms, and diacritics.Source: §5.1 · p. 22, §5 · p. 22

5.2. Potential Applications

  • The dataset is intended for training and evaluating line-level OCR and can support research on mixed-direction recognition, script-aware OCR, multilingual sequence modeling, and rendering robustness.Metadata can be used to construct controlled subsets for analyzing specific OCR challenges.Source: §5.2 · p. 23

5.3. Limitations and Future Work

  • Because MuSTEL is synthetic, it cannot reproduce every real-document characteristic, and its line-level scope excludes page layouts, text detection, and other document-structure problems.Real documents may include physical damage, handwriting, complex scanning artifacts, or layout interactions that are not fully represented.Source: §5.3 · p. 23

6. Conclusion

MuSTEL is a large-scale synthetic dataset for Persian, Arabic, English, and mixed-language line-level OCR, designed to address gaps in multilingual coverage, typographic diversity, and document-like visual conditions. Its statistical analysis supports its use in multilingual and mixed-direction recognition research while noting limits in representing real documents and page layouts.

6. Conclusion

  • MuSTEL contains 14,527,988 annotated text-line images generated from multilingual corpora, organized into a large training set and two predefined evaluation sets.Its validated font repository contains 2,009 font files from 711 families, alongside varied image dimensions, backgrounds, and rendering conditions.Source: §6 · p. 24
  • The framework extends TRDG with script-aware processing, bidirectional composition, capability-based font selection, multiple-font and style rendering, and automated annotation and quality control.The dataset also includes substantial lexical and character-sequence diversity.Source: §6 · p. 24
  • MuSTEL's linguistic, typographic, and visual diversity makes it suitable for multilingual and mixed-direction line recognition research, particularly in document-oriented OCR pipelines.Statistical analysis finds broad coverage of Persian, Arabic, and English, plus a substantial number of mixed-language samples.Source: §6 · p. 24
  • Because MuSTEL is synthetic and line-level, it does not fully reproduce real scanned or photographed documents or complete page layouts.The paper identifies transfer to real-world OCR data, evaluation with modern recognition architectures, and extensions to more languages, visual conditions, and page-level resources as future work.Source: §6 · p. 24
Loading 2610.34148v1…