Building a Naive OCR System
This document summarizes a student project to build a naïve optical character recognition (OCR) system. The students limited their system to one font from one book, lowercase characters only, and pictures taken with one phone. They collected data from 44 pictures containing 242 words. Their workflow involved preprocessing images then extracting 8x16 pixel feature vectors for each character. They used a support vector machine classifier with a radial basis function kernel, achieving 98.86% accuracy on 10-fold cross-validation of their data, compared to 91.47% for commercial Tesseract OCR. Key lessons were that character frequencies vary significantly from English, not all fonts separate easily into characters, and installing libraries was more difficult than expected.
Building a Naive OCR System
- 1. BUILDING A NAÏVE OCR SYSTEM Kaur Alasoo and Jaana Metsamaa Data Mining (2009 fall) 28 January 2010 http://kaur.pri.ee/projects/wordocr/
- 2. VISION Take a snapshot of a page.
- 3. VISION Extract one word from the picture.
- 4. VISION Look up the word on Google, dictionary, etc...
- 5. REALITY • To increase the likelihood of success, we limited ourselves to: • One font from one book. • Only lowercase characters. • Pictures taken with one phone.
- 6. GETTING THE DATA 44 pictures
- 7. GETTING THE DATA 44 pictures 242 words
- 8. 0. Take a picture 1. Convert to grayscale 2. Apply thresholding WORKFLOW 3. Cut horizontally 4. Cut vertically 5. Resize 8x16 px
- 9. FROM PICTURE TO FEATURE VECTOR
- 10. FROM PICTURE TO FEATURE VECTOR
- 11. FROM PICTURE TO FEATURE VECTOR 255
- 12. FROM PICTURE TO FEATURE VECTOR 255 255
- 13. FROM PICTURE TO FEATURE VECTOR 255 255 255
- 14. FROM PICTURE TO FEATURE VECTOR 255 255 255 0
- 15. FROM PICTURE TO FEATURE VECTOR 255 255 255 0 0
- 16. FROM PICTURE TO FEATURE VECTOR 255 255 255 0 0 155
- 17. FROM PICTURE TO FEATURE VECTOR 255 255 255 0 0 155 255
- 18. FROM PICTURE TO FEATURE VECTOR 255 255 255 0 0 155 255 255
- 19. FROM PICTURE TO FEATURE VECTOR 255 255 255 0 0 155 255 255
- 20. TRAINING THE CLASSIFIER We used 1343 characters to train a SVM with RBF kernel.
- 21. RESULTS • 10-fold cross-validation accuracy with our method: 98.86% • Character classification accuracy using Tesseract OCR: 91.47%
- 22. WHAT WE LEARNED?
- 23. SOME LETTERS ARE EXTREMELY RARE 15 English Our data 11 8 4 0 e t a o i n s h r d l c u m w f g y p b v k j x q z
- 24. NOT ALL FONTS CAN BE EASILY SEPARATED INTO CHARACTERS
- 25. WHAT WE LEARNED? • Installing libraries is the most difficult thing. • Do not overly restrict yourself. • Donot build your on OCR system unless its absolutely necessary. • The letters are the easy part. http://kaur.pri.ee/projects/wordocr/