OCR Text Reader Comparison

Overview:

In this project, I evaluated and compared three popular Optical Character Recognition (OCR) libraries—Textract, EasyOCR, and Tesseract—by using them “out of the box” to analyze a diverse set of images containing text. My goal was to understand how well each tool performs without extensive customization or preprocessing.

GitHub: https://github.com/BrandtBrandtBrandt/OCR_Text_Reader_Comparison.git

Tools and Setup:

• OCR Libraries: Textract, EasyOCR, Tesseract

• Platform: Google Colab

• Dataset: A collection of 100 diverse photos containing text, kindly labeled by Felipe from Computer Vision Engineer.

Evaluation Metric:

I used the Jaccard Index to measure how many words each model correctly identified from the dataset. The results for each library were as follows:

• Tesseract: 0.0117

• EasyOCR: 0.2228

• Textract: 0.3445

Results and Insights:

While Textract emerged as the most accurate library with a score of 34% of words correctly identified, it is clear that even more work would be required for production-ready accuracy. Tesseract performed poorly in this setup, likely due to its reliance on preprocessing steps to optimize text detection. This comparison underscores the importance of preprocessing for Tesseract and highlights Textract’s relative strength for “out of the box” performance.

Acknowledgment:

This project was part of the Computer Vision in 30 Days Coding Challenge by Felipe, also known as Computer Vision Engineer on YouTube.