Best approach to detect read numbers from 1 to 99

edit

Hi !

I'm currently using OpenCV to detect things on a video and those things have sometime a number printed on, which I want to read. Theses numbers could be from 1 to 99 and I'm able to extract them in a binary image as :

http://vps166675.ovh.net/img126.jpg.b... http://vps166675.ovh.net/img130.jpg.b... http://vps166675.ovh.net/img101.jpg.b...

I try these approches : - Cascade classification using HAAR and HOG classifier : very poor results even if I work with ~500 positives images and thousands of negatives (I train these classifier using gray-scale images) - Tesseract OCR, but, even with these binary images, and even if I told him that there is only digits, it reads 3 as many as 8, so it is not reliable (And it actually reads nothing most of the time).

Have you any advice to handle these case ? Maybe a feature type that I don't try ? A hack using hough line detection or things like that ?

Thank you in advance.

Damien.