issai/Wild1024

Multilingual Wild Font Identification Dataset:

Dataset Summary

This dataset consists of highly diverse cropped text images extracted from documents in the wild, natural scene images, book covers, signs, handwritten text, and various other sources, categorized by language and font family. The primary goal of the dataset is to facilitate research: Font Identification (Font ID) and multilingual Optical Character Recognition (OCR) under realistic, non-ideal conditions.

The corpus contains image patches of text, natively labeled with their corresponding language script and typographic font family. It is intended primarily for research on typography analysis, domain adaptation (synthetic vs wild), and evaluating OCR robustness across varying font families.

Dataset Structure

Each example in the dataset contains the following fields:

Feature	Description
image	The cropped image segment containing the text.
lang	The language of the text in the crop (e.g., arabic).
font_family	The specific font family or typographic classification (e.g., sans-serif).

Dataset Statistics

• Total Images: 1,024
• Number of Languages: 11
• Number of Font Families: 4

Images Per Language

• arabic: 93
• azerbaijani: 93
• uzbek: 93
• russian: 93
• french: 93
• spanish: 93
• turkish: 93
• english: 93
• chinese: 93
• kazakh: 94
• kyrgyz: 93

Images Per Font Family

• serif: 267
• sans-serif: 264
• handwritten: 262
• monospaced: 231

The dataset was collected and curated by the authors from publicly available sources, as well as from text materials generated by the authors. It contains only isolated, cropped text segments, with the broader document context intentionally removed. The dataset is intended strictly for non-commercial academic use and research in document analysis.