-1
$\begingroup$

I want to detect textures on images, for example after proccessing on image must be shown areas of each texture and area where is no texture.And I want to compute "mean" texture of selected area.

how can I do that?

Update: yes, I know that I can use some corelation technic with some small patches or just split whole image in some small areas(for example with superpixel) and computes some descriptor of area and then cluster them.But I think problem is that texture patch may be different in size.And task seems computional huge,maybe there is some faster FFT approach? And is there any method to determine is area texture or not?


http://masters.donntu.edu.ua/2006/kita/varshavskaya/library/art07.htm I found an article about.But still loking for more examples.

$\endgroup$
2
$\begingroup$

One way to do that is to use template matching.

The template matching can work based on image data - i.e. compute correlation between sample texture patch and the image.

Another way is to use feature-based method, which is useful in case the texture is geometrically distorted. This method requires extracting features from sample texture patch and from image and assign a descriptor vector to each feature. The two sets of features are then matched using their descriptors in a nearest-neighbor fashion.

Yet another approach can be active contours or graphcut based methods, but I am afraid these require some initialization, like marking of the texture areas to fill them up.

$\endgroup$
  • $\begingroup$ yes, I know that I can use some corelation technic with some small patches or just split whole image in some small areas(for example with superpixel) and computes some descriptor of area and then cluster them.But I think problem is that texture patch may be different in size.And task seems computional huge,maybe there is some faster FFT approach? And is there any method to determine is area texture or not? $\endgroup$ – mrgloom Aug 15 '12 at 7:08
  • $\begingroup$ I have too little knowledge about template matching, but feature-based method solve the scaling problem as the features can be scale invariant (descriptors are always computed on appropriate scale). Maybe wavelets can deal with the scaling problem. $\endgroup$ – Libor Aug 16 '12 at 14:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.