Master LLMs with our FREE course in collaboration with Activeloop & Intel Disruptor Initiative. Join now!


How To Make a Synthesized Dataset To Fine-Tune Your Ocr
Data Science   Latest   Machine Learning

How To Make a Synthesized Dataset To Fine-Tune Your Ocr

Last Updated on January 29, 2024 by Editorial Team

Author(s): Eivind Kjosbakken

Originally published on Towards AI.

Fine-tuning your OCR engine to your specific use case is required if you want to achieve state-of-the-art performance for your situation. Fine-tuning an OCR engine requires a large dataset, however, which is expensive to annotate. Instead, you can create your own fully synthetic dataset, requiring no annotating, which I will show you how to do in this article.

Learn to create a fully synthetic receipt dataset with this tutorial. OpenAI. (2024). ChatGPT [Large language model]. /g/g-2fkFE8rbu-dall-e

· Motivation· Making a baseline· Finding the bounding boxes in the original image· Loading the annotated data· Adding word labels· Creating synthetic receipts· Adding variation to the receipts ∘ Augmentation 1: Add brightness ∘ Augmentation 2: Add noise ∘ Augmentation 3: Add perspective change ∘ Augmentation 4: Add sharpening ∘ Augmentation 5: Add Gaussian blur ∘ Augmentation 6: Add contrast ∘ Augmentation 7: Add saturation· Apply the augmentations· Future work· Conclusion· References

My motivation for this article is the fact that I tried applying EasyOCR to read the text off some Norwegian receipts, but I found the results not to be satisfactory. The Levenshtein distance achieved on the off-the-shelf EasyOCR was around 75%, which is not good enough when it comes to extracting information from supermarket receipts…. Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI

Feedback ↓