caveman273/aida-handwritten

Handwritten OCR training data from AIDA-project

Dataset Summary

This dataset contains handwritten textline images and their transcriptions from the AIDA-project. It is a subset of the full AIDA dataset, containing only the best-quality handwritten annotations — lines where the annotator was confident about every character. The majority of lines are in Finnish, with some Swedish, English, French, and German.

Supported Tasks

The dataset was created for handwritten text recognition (HTR).

Languages

The majority of the textlines are in Finnish, but some are in Swedish and English. In addition to this there are few French and German textlines.

Dataset structure

Data Instances

Each row contains:

• image: the textline image (PNG bytes + filename)
• text: the transcription
• file_name: the original image filename

Data Fields

Field	Type	Description
image	Image	Textline image
text	string	Ground-truth transcription
file_name	string	Original image filename

Data Splits

This dataset contains only the "best" handwritten annotations (every character understood by the annotator). The number in parentheses shows the additional "semi" lines (some characters unclear) not included here.

Dataset Split	Handwritten
Train	6943
Validation	1151
Test	1270

Dataset Creation

Source Data

The data is collected from Central Archives for Finnish Business (ELKA). It consists of various document types including letters, ship records, business publications etc. It includes correspondence between companies, organizations and the public.

Who are the source language producers?

Given the various types of archival material used in annotation, the scope of producers of the original texts is broad. It includes private individuals and employees of different companies.

Annotations

The textlines were first cropped out of the original image and then transcribed. If the transcription was unclear, the annotator marked it as either "somewhat unclear" or "unclear". Unclear images were discarded, but the "somewhat discarded" images are presented here as in the "semi" annotation files. The rough estimate for "somewhat unclear" class is that less than 100% and more than 50% of the characters are unclear.

Who are the annotators?

Annotators were employees of National Archives of Finland and ELKA.

Synthetic data

As a way to increase the amount of training data, we created synthetic data by using this library https://github.com/Belval/TextRecognitionDataGenerator. We collected Finnish books from https://www.gutenberg.org/ and Finnish magazines from https://archive.org/ and created different kinds of textlines. The different kinds include normal textlines, rotated textlines, textlines following a sinosoidal curve and textlines where characters are subjected to noise.

Personal and Sensitive Information

The dataset is not anonymized, so individuals' names can be found in the dataset.