Topic: Image processing (Page 1 of 1)

I've actually got two areas of interest here...

1) Pattern matching: I have an image of some large object and I want to locate and isolate a particular area of it (a fairly small area) based upon certain geometric features in the general area I'm interested in.
Is 'pattern matching' the correct term for this even? Are there other names or descriptions for this that I can use when searching for information?
Does anyone know of (or used) any utilities that would accomplish this?


2) Digital character recognition: I'll have an image that has an alphanumeric string in it and I'd like to be able to read the string programmatically. Something along the lines of the way WhatTheFont works, except I won't care about the font, just the text.
I've found a few projects that seem to be aimed at this:
OCRE
JOCR
http://sourceforge.net/projects/oocr]OOCR
Has anyone worked with anything like this? Any recommendations of what to look for or sources of information?


For both of these, I'm just looking for a proof of concept right now, so:
- trial/demo is good
- free is excellent
- open source is better


And as always, any information and/or resources are appreciated. I haven't found a whole lot on pattern matching (very possibly because I'm looking for the wrong thing), but I've found tons on DCR/OCR and I need to separate the useful information from the common, not-so-useful information

Many thanks,
LF

--

Any sufficiently advanced bug is indistinguishable from a feature.

I think it's pattern recognition you're after...as explained here.
http://www.answers.com/main/ntquery?s=pattern%20matching&gwp=16

Just in case you don't dig far enough. =)

http://www.answers.com/topic/image-analysis

http://www.answers.com/topic/facial-recognition-system

___________________________________________________________________________
The goal in Life's Journey is not to arrive at the grave safely in a well preserved body, but rather to skid in sideways, totally worn out, shouting "holy moly what a ride!"

(Edited by NoJive on 08-21-2006 18:15)

I've tried GOCR and Ocrad. I can't remember which, but one was far better than the other. I think it was the one that uses xpm images and outputs straight to stdout.

NoJive: Thank you, I think 'pattern recognition' is closer to what I'm picturing.

I'm not sure it's exactly it though... Most of the resources I've found that talk about it have much more advanced and complicated "patterns" that they are looking for. All I need is to take an image of a big object and focus in on a smaller piece of it.
Let's say I have a picture of a large rectangle. Somewhere inside the rectangle there is a circle and a triangle (these will always be the same size and close to the same orientation, within 5-10% of their dimensions), and I want to isolate the section of the larger image containing the circle and triangle.

Does this still sound like 'pattern recognition'? Am I going nuts? Should I just beat my head on my desk for awhile, go home, and start again tomorrow?

resio: I'd found GOCR but haven't had a chance to look at it yet. Ocrad I hadn't found, thank you.

--

Any sufficiently advanced bug is indistinguishable from a feature.

http://sourceforge.net/projects/tesseract-ocr

Just ran across this one, haven't tried it.

FYI: I read that it only does English and took ten years to make, so it's probably a total piece of shit.

Pattern recognition [b]Algorithm(s)[b] should get you to loads of theory, but theory that can be used directly.
The terms are right.

Since pr is being researched at a very advanced AI level, you may stumble into much more than what you want.
I which case I recommend reverting to - basic brain functions and logics: if color is the main difference between your shapes,
then you can build an algorithm easilly along the lines of "planar subdivision" - subdivide the picture into small squares,
and check each square's color uniformity - just a thought.

Lord_Fukutoku: Actually those things were a big part of my BS course.

Could you post some sample images so we can see if you have to use the heavy artillery or a straight forward technique.

First of all, all this is called Digital image processing.

You should search for image segmentation. This term cover the image processing techniques used to separate different regions of interrest in an image. On wikipedia the french page seems better than the english one.

For the characters/digits recognition, you might need to generate the skeleton ( see topological skeleton ) of the regions ( obtained above ) and check the shape of the skeletons and see if they match some reference skeletons. If you aim for a pixel based approach, you might resize the different regions to a given resolution and compare them to the template of the patterns you are looking for. *hint* blurring both the template and the regions improves the tolerance of the algorithm.

Reading about OCR might help too. And in that field CAPTCHA defeating technique should give some precious tips.

One thing we did during my BS was to make an FFT on the pattern we wanted to find in another image, then use the resulting images ( in the frequency domain ) as a kernel we applied on the FFT of the image supposedly containing the pattern. Doing an IFFT on the convoluted image gave a image that was mostly black, with some white pixels exactly where the pattern appeared in the initial image. I don't remember the gore(?) details but that's basically the idea. [edit] This is called 'Shape recognition by convolution' [/edit]



(Edited by poi on 09-02-2006 00:29)

If it is OCR that you are needing, Slashdot is running an article on Tesseract OCR. The article says that

quote:

---

it has been touted as one of the most accurate Optical Character Recognition (OCR) programs available.

---

You can download it here

Later,

C:\

quote:

---

Since pr is being researched at a very advanced AI level, you may stumble into much more than what you want.

---

Yep, so far that's been the case.

quote:

---

Could you post some sample images so we can see if you have to use the heavy artillery or a straight forward technique.

---

Unfortunately I don't have any images yet that I would be using in a real case, however, the closest example I can think of right now is to take a picture of the front (or back) of a car and isolate the license plate (assuming you know the color and size of the plate (since it varies from state to state and country to country).
Thank you for the other pointers too poi.

C:\ & reisio - Thanks for the link, I hadn't seen that one yet.


Thanks everyone for the tips and pointers everyone, I really appreciate it. Work on this has kinda slowed down for the moment though because of testing at work and I just started a compilers class at the univ (whoo ). Once I get back to this project and process the new info, I'm sure I'll be back with some more questions.

--

Any sufficiently advanced bug is indistinguishable from a feature.

(Edited by Lord_Fukutoku on 09-06-2006 15:20)