1
$\begingroup$

In this paper under section 2. "Proposed Methodology", the author creates a GLCM from an image and then extracts texture features resulting in a new image for each feature. I've seen other authors do this in other papers as well. The texture features used in the image are the common Haralick features detailed on page 619 here.

  • How did the author extract a texture feature image from those features?

    The author has a unique image for contrast, an image for entropy, an image for correlation, etc. When I look at the Haralick feature equations the result is always a number not another image.

  • How is a unique image created for each texture feature?

EDIT:

For GLCM definition as suggested in comments, a GLCM stands for the Gray-level co-occurence matrix. It's a common technique used for texture analysis. It is defined over an image to be the distribution of co-occurring pixel values at a given offset.

$\endgroup$
  • $\begingroup$ Hi! Welcome to DSP.SE! Could you be as nice as to (very briefly) define what GLCM is? I suspect I won't be able to answer, but maybe it makes your question easier for people who are knowledgeable about image analysis but don't know GLCM. $\endgroup$ – Marcus Müller Jan 17 '17 at 18:08
1
$\begingroup$

Looking at these references, this is what I would understand:

  1. The Haralick descriptors only create a single value for each GLCM (as you also thought).
  2. The Rampun paper creates 32 images, describing 32 different features for each input image.
  3. The crucial point is this sentence: On the other hand, we used a small window size of 5 x 5 throughout the process This sentence, in combination with this one (taken from here, p.25, Sec. 2.6): The GLCM Texture Image is the result of moving the GLCM window across the entire image or sub-image.

So, my understanding is:

  • Create moving window of size 5x5
  • move this window over your input image
  • for each window position (x,y), calculate a feature (e.g. contrast), and write the value for this feature into the "contrast" image at position (x,y).

This way, you extract "local" features, i.e. the GLCM is only done over local area of size 5x5. This also makes more sence, since the Rampun paper wants to segment images, i.e. it needs local features instead of global ones.

$\endgroup$
  • $\begingroup$ Thank you. At one point I had thought this but dismissed it because I felt a 5x5 would be too small to extract relevant features. But upon reading the section you pointed to on pg. 25 it is clear that you are correct. I've been stuck on this for a very long time. Thanks again for the help. $\endgroup$ – user2012741 Jan 17 '17 at 19:39

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.