Texture analysis for a large image size

Hello, I am trying to perform texture analysis on a large grayscale images (with high resolution) trying to find statistics derived from GLCM and other histogram properties(e.g skewness) which I intend to use as an input for classification solver e.g (ANN/SVM) to find small features properties in the texture in the image (e.g . grain size distribution). the question is do I use greycomatrix() for the whole image or shall I using a moving window frame to find GLCM properties for each pixel and then compute the mean/std/var/min/max for the GLCM statistics (contrast/energy ..etc) matrices matrix.

If I need to use a moving frame how do determine the optimum size for example shall I use 5*5 or 15*15 window size.

Thanks.

Answer 1

[Image Analyst]

If you're trying to find regions of the image that have a certain texture, then you can't just compute the GLCM of the whole image because that won't give you any local information. So you'll have to have a scanning window. Take a look at nlfilter() or blockproc(). Experiment with different window sizes that are a percentage of the image to find the percentage that works well for most of the images you expect to encounter.

Answer 2

[Constantino Carlos Reyes-Aldasoro]

As mentioned previously, the co-occurrence matrix will compute the occurrence of pairs of values at different angles and different distances, and extract a metric (like the mean or variance) and then add all values. This is useful if you want to classify one image, but if what you want to do is partition into regions (segmentation by texture), what you need is to obtain those metrics per-region, and then classify. Have a look at this detailed guide of the use of the co-occurrence matrix:

http://wrap.warwick.ac.uk/61393/

Finally, GLCM is not the only way to analyse texture, it is easy and simple, but rather time consuming and even back in 1999 there were better ways to segment by texture. Have a look at this repository which uses sub-band filtering for texture segmentation

https://uk.mathworks.com/matlabcentral/fileexchange/67931-reyesaldasoro-texture-segmentation