IAM Handwriting dataset

Dataset Overview

Dataset Name

IAM Handwriting dataset

Dataset Contents

• Includes 115,320 isolated, labeled word images.
• Written by 657 distinct authors.

Dataset Download

• The dataset can be downloaded here.

Intended Use

Used for handwritten text recognition, employing convolutional neural networks (CNN) and bidirectional GRU (Bi‑directional GRU) with CTC decoding.

Performance

• Recognition accuracy on the test set is 59%.
• Errors may stem from improper handling of GRU gates.

Future Improvements

• Plans to utilize cloud virtual machines and pre‑trained language models to correct spelling errors and boost recognition accuracy.