Jump to content
  • Advertisement

Archived

This topic is now archived and is closed to further replies.

_nomad_

On image comparison

This topic is 5211 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

say i have two images, one is a face (with eyes, nose, and mouth), and another is an eye image. i want the software to auto-find where in the face image, the eye image fits...any ideas on this? per-pixel won''t be good, since the eye can also be partially covered with some strands of hair, for example... thanks.

Share this post


Link to post
Share on other sites
Advertisement
Depending on the complexity of the matching you could use anything from a simple difference scheme where you look at how different the particular pixels are all the way up to using a learning neural networking.

Share this post


Link to post
Share on other sites
let's assume that the eye image and the eye in the face image both has exactly the same colors...except that the eye in the face image might have some strands of hairs covering it....how to go about auto-detecting the x/y location of the eye in the face?

what's the simplest working scheme would you use? and links to some tutorials would be great.

thanks.

[edited by - _nomad_ on June 3, 2004 11:10:49 PM]

Share this post


Link to post
Share on other sites
no tutorials needed, it''s a simple search

think about this example sentence.

now think if i wanted to find about in that sentence

naive approach: you look for the first a, then you test to see if bout follow.

I''ll post an improved approach later.. when I have more time.

but that''s the basic idea, you can iterate through it, and then once you have it working, improve on your implementation.

at that stage, you can probably just search up some text searching techniques.. and apply them to images (they''re both just data)

Share this post


Link to post
Share on other sites
searching per letter would be like searching per pixel...however, i can''t serch per pixel since different images have different small entities obscuring the query image...like in the eye on the face image, i could have 50 faces and 50 pairs of eyes, each one have different types of hairs and different number of hair strands covering their eyes...

help!

thanks.

Share this post


Link to post
Share on other sites
google pattern recognotion and analysis and image recognition and analysis. check the spelling of recognition too

what you''re trying to do still hasn''t been "solved" - there are hundreds of different ways of doing it, but most of them revolved around trying to get an "almost perfect" fit.

take the example with the eyes and the hairs. The eyes would match (giving it a perfect score) but the hairs would not (lowering the score slightly). You will have to allow for less than 100% matches otherwise you''ll never pick up anything. the challenge lies in creating an algorithm capable of dealing with this.

Share this post


Link to post
Share on other sites
Guest Anonymous Poster
You should have a look at FFT ( Fast Fourrier Transform )...

Euh... I''m not sure, but my understanding is that it can apply in this case. There was an article on image processing on Gamedev dealing with this.

Hope it helps,

Share this post


Link to post
Share on other sites
i tried histogram comparison...but it doesn''t work well since i have to shrink each pixel''s bits (24bit pixels would need an array of size 16.7million)...

...i''ll try and find the FFT article here on gamedev...thanks.

Share this post


Link to post
Share on other sites
its not a simple task, i think you can find the eye as follows:
1st you will have to get outlines (FFT and a high-pass filter should do that)
2nd use Hough-Transformation for spheres/ellipses on your outlines to find the iris-outlines (at least some candidates), you can check for structures where one spheres/ellipse is enclosed by another
3rd now that you have a possible candidate transform this imagepart into a normalized form (so your ellipse will be spherical)
4th compare, rotate, compare..., or use a Hough-Transformation (this time on lines) and compare the structure you get with the Hough-Transformation on your referenzimage to find a good rotation, and compare then

but i have to tell you: iam not 100% sure that this will work...
and a Hough-Transformation is a very handy way to get Geometrical objects out of points (google it)... but iam not sure how it works with spherical/elliptical objects... the class i visit had only Hough Transformations for lines (but my prof said, there are equivalent ones for spheres etc...)


hope that helps
T2k

Share this post


Link to post
Share on other sites
As people have already said, you''d probably find yourself needing to break the images down into more high-level constructs - curves, ''color zones,'' and so on.

Then you''d look to compare the two, and calculate a percentage match (i.e. based on how close it is).

Then you''d pick a threshold value - say 98% - and call anything above that a match, and anything below it a fail.

Share this post


Link to post
Share on other sites

  • Advertisement
×

Important Information

By using GameDev.net, you agree to our community Guidelines, Terms of Use, and Privacy Policy.

We are the game development community.

Whether you are an indie, hobbyist, AAA developer, or just trying to learn, GameDev.net is the place for you to learn, share, and connect with the games industry. Learn more About Us or sign up!

Sign me up!