Results 1 to 2 of 2
- 03-08-2009, 05:04 PM #1
how to use wavelets instead of FFT's for ocr
In trying to distinguish various real-world patterns, OCR fails readily on several matters that are near-ovious for the human mind. Wavelets are considered to be a potenetial improvement on FFT for some of these issues. ( @todaly ) What uses could wavelets be applied to in the problem domain and how would one code that in Java?Introduction to Programming Using Java.
Cybercartography: A new theoretical construct proposed by D.R. Fraser Taylor
- 03-10-2009, 06:50 AM #2Senior Member
- Join Date
- Jan 2009
- Rep Power
As far as the 'coding it' part goes, I have no idea. I've done a LOT with FFTs and only dabbled in wavlets. However, the kinds of problems you've been discussing (localized pattern recognition where scales can shift) is what wavelets were invented for. Wavelet - Wikipedia, the free encyclopedia
Ok, so wavelets are good for localized analysis with scaling (which is what OCR is 50% about).
The other 50% is rotation. If you scan something, it's going to be off normal by some amount. I would probably try to histogram the pixels to figure out rotation instead of applying a neural network (the efficient general purpose solution) or a brute force method (the easy solution).
(Since everything rotates together, some simple statistical approach will probably work).
If I were given you're assignment, I would have to start reading up, or contract an OCR expert to help out.