Telugu Document Image Dataset

Key Features:

1. Wide Variety of Realistic Scans: The dataset includes images mimicking realistic variations, such as aging effects, smudges, and incomplete characters, commonly found in physical book scans or older documents.
2. Image-Text Pairing: Each image is linked with its corresponding pure text sequence, allowing models to learn efficient text extraction without additional manual preprocessing steps.
3. Customizable Data Generation: The dataset is built using open-source generator code, which provides flexibility for users to adjust hundreds of parameters. It supports custom corpora, so users can replace the probabilistically generated “gibberish” text with actual texts relevant to their use cases.
4. Scalable and Efficient: Thanks to parallelized processing, larger datasets can be generated rapidly. Users with powerful computational resources can expand the dataset size to hundreds of thousands or even millions of pairs, making it an adaptable resource for large-scale AI training.
5. Multi-Script Support: The generator code can easily be extended to other scripts, including different Indic languages or even non-Indic languages, by modifying the Unicode character set and adjusting parameters such as sentence structure and paragraph lengths.

Dataset Applications:

This dataset is especially useful for developing OCR systems that handle Telugu language documents. However, the data generation process is flexible enough to extend to other Indic languages and non-Indic scripts, making it a versatile resource for cross-lingual and multi-modal research in text extraction, document understanding, and AI-driven translation.