Dataset Viewer
Auto-converted to Parquet Duplicate
image
imagewidth (px)
264
2.33k
json
stringlengths
474
217k
{"page_information":{"width":794,"height":1123,"dpi":150,"image_width":1557,"image_height":2203,"id":"000000","seed":"205677028794804-0","format":"a4","template":"flow","margin":{"vertical":55,"horizontal":59},"font_scale":1.0,"masthead":false,"columns":2,"flow":"ltr","elements":16,"background_image":"coco_000000557884...
{"page_information":{"width":816,"height":1056,"dpi":100,"image_width":850,"image_height":1100,"id":"000001","seed":"205677028794804-1","format":"letter","template":"flow","margin":{"vertical":66,"horizontal":85},"font_scale":1.0,"masthead":false,"columns":1,"flow":"ltr","elements":14,"background_image":"coco_000000322...
{"page_information":{"width":794,"height":1123,"dpi":148,"image_width":1224,"image_height":1731,"id":"000002","seed":"205677028794804-2","format":"a4","template":"paper_figure_page","margin":{"vertical":53,"horizontal":45},"font_scale":1.0,"masthead":false,"columns":1,"flow":"ltr","elements":6,"background_image":null,"...
"{\"page_information\":{\"width\":816,\"height\":1056,\"dpi\":98,\"image_width\":1230,\"image_height(...TRUNCATED)
"{\"page_information\":{\"width\":794,\"height\":1123,\"dpi\":128,\"image_width\":1059,\"image_heigh(...TRUNCATED)
"{\"page_information\":{\"width\":816,\"height\":1056,\"dpi\":107,\"image_width\":937,\"image_height(...TRUNCATED)
"{\"page_information\":{\"width\":794,\"height\":1123,\"dpi\":147,\"image_width\":1216,\"image_heigh(...TRUNCATED)
"{\"page_information\":{\"width\":794,\"height\":1123,\"dpi\":131,\"image_width\":1416,\"image_heigh(...TRUNCATED)
"{\"page_information\":{\"width\":794,\"height\":1123,\"dpi\":137,\"image_width\":1133,\"image_heigh(...TRUNCATED)
"{\"page_information\":{\"width\":794,\"height\":1123,\"dpi\":125,\"image_width\":1034,\"image_heigh(...TRUNCATED)
End of preview. Expand in Data Studio

Synthetic OCR pages

32,000 synthetic document pages in Arabic, English and Chinese, rendered from generated HTML in Chromium, with exact word, line, paragraph and element boxes.

Each row has two columns:

  • image: the page image (PNG).
  • json: the page's annotation as a JSON string: page_information (format, template, size, DPI, seed, photo or scan effects), word_bboxes, line_bboxes, para_bboxes, element_bboxes and relation_graph.

Load one subset, or all:

import json
from datasets import load_dataset

pages = load_dataset("FaisalHejary/synth_OCR", "newspapers", split="train")
annotation = json.loads(pages[0]["json"])

Subsets

Subset Pages What it covers
general_mixed 2,560 Every format and template at the generator's mix (flow at most 20% of each format), the baseline the targeted sub-pipelines add to; rich inline formatting; a third scanned.
ar_small_text 2,816 Arabic-only body text at 7-16 image px (low dpi, dense pages), figures in running text, print spelling variants, bold phrases; a fifth of the pages scanned at high scan resolution (so the text stays legible); a tenth of the text fully vocalized.
hard_scripts 1,536 Body text and headings in the families OvisOCR3 reads worst (Ruqaa, Nastaliq, Kufi, Mirza; Chinese grass-script and brush faces), at normal and small sizes, with tables and vocalized text; a fifth scanned, some photographed.
ar_numeric_tables 2,560 Arabic-only pages of tables: generated figure tables (Arabic-Indic digits, Arabic and ASCII separators, signed values in every sign convention) beside real Arabic document and World Bank tables, grouped two-level headers, merged and blank cells; figures in running text.
receipts_till_ar 1,024 Arabic till receipts (thermal, restaurant, grocery): sums in Arabic-Indic digits and Arabic separators, negative discounts in every sign style, 0-3 decimals, product and cashier names; often photographed or scanned.
receipts_till_mixed 768 Till receipts in Arabic, English and Chinese with product and cashier names and 0-3 decimal currencies (Arabic sums on the Arabic ones), often photographed or scanned.
receipts_invoices 1,024 Full-page invoices and two receipts scanned side by side (A4, letter), with Arabic sums on the Arabic ones and figures in Arabic running text.
newspapers 2,560 4-6 column newspapers, mostly Arabic, with running headers, page numbers, foreign-language items, datelines and long titles that can sit at the top left, inline code-switching, print spelling variants, figures in running text and Kufi or Ruqaa display headings.
ar_multicolumn_layouts 1,920 Arabic magazine, changing-layout, sidebar and panel pages in 2-3 columns with spanning titles, figures and Latin or neutral items; plain 3- and 2-column RTL pages; and RTL column pages whose left column opens with a key figure or a heading (columns_corner); long titles (7-11 words), datelines and URLs as headings, Kufi or Ruqaa display headings.
books_poems 1,792 Arabic book pages (matn and sharh, hashiya, frames) with one or two long two-hemistich poems (16 verses or more, most drawn with 8 or more; 1-3 on A4 pages) and poem pages, classical and fully vocalized text; scans down to 0.55 of the size (the text is large).
academic_papers 1,792 Paper pages in the 14 academic layouts with display formulas, tables, figures, heading levels and references, in all three languages.
reports_charts 1,280 Report layouts: key figures, dashboards, tables (generated and real Arabic ones) and figures, with figures in running text.
slides 1,792 All 20 slide layouts: dark pages, background shapes and pictures, key figures, every font family; scans down to half resolution.
cv_forms 1,280 CV layouts, memos, circulars and question-answer pages: labels beside values, dates, rich inline formatting.
code_switching 1,536 Pages that switch language element by element (60%) and inside sentences and single lines (half the passages, headings, list items, captions and quotes): Arabic beside English and Chinese words.
scans_rescans 2,560 Every format scanned (90%), most of them rescanned up to three times, down to 0.75 of the page's size a pass, with black-and-white scans, JPEG, small-print Arabic pages and vocalized text; no watermarks or dot-only spelling variants (a binarized rescan erases them).
photos_captured 768 Every format photographed (95%): turned up to 10 degrees, perspective, blur, noise, cast shadows, folds and stains, with room around the page so the turn is not clamped into a shear.
photos_rotated 512 Pages photographed upside down (172-180 degrees), blurred and noisy (blur bounded so the text stays legible), at low dpi.
furniture_watermarks 640 Running headers, page numbers, footnotes and one to three watermarks on almost every page.
cjk_latin 1,280 Chinese and English pages: Chinese newspapers, small Chinese print, vertical text, spaced and uppercase Latin headings, drop caps; every font.
Downloads last month
110