Saturday, April 4, 2015

OCR Demo

 Here is a video with a little OCR demo. The demo show the performance of an OCR engine I have implemented. This OCR engine is designed for very small font size. Normally, most publicly-available OCR engines are designed for (paper) document scans. The size of text in these kind of scans (images) is at least twenty pixels high. For example, Tesseract, probably the most popular free OCR engine, is not able to read text (font size) less than twenty pixels high.

The main challenge with very small text is that the resolution is very low. As a result, it is common that different letters connect and merge with each other. So, the main problem becomes to segment the individual letters.

Very-low resolution fonts are typical of computer images. The task of reading (OCR) a computer image is, on one hand, easier than OCR of a (paper) document scan, in that the variations in how each letter of the alphabet looks like are much more limited: a specific letter in a given font style, will always look the same. However, OCR of computer images, on the other hand, present a bigger challenge in terms of segmenting the individual letters.

The following OCR demos deals with the added difficulty that there is text in different colors. The complexity of variations in ink colors is that letter segmentation becomes even harder. In addition, the individual letters do no longer look the same on different colors (even the font style remains the same).

OCR Demo

Tuesday, October 14, 2014

Author: Javier Bautista
Skype: jbscheuber

Computer Image 1



Computer Image 1 - Probability Map


Computer Image 2


Computer Image 2 - Probability Map


Computer Image 3


Computer Image 3 - Probability Map


Noisy Image 1


Noisy Image 1 - Probability Map


Newspaper Image 1


Newspaper Image 1 - Probability Map
 

 Newspaper Image 2


Newspaper Image 2 - Probability Map


Newspaper Image 1 - Text Detection


Invoice Image 1 - Text Detection