Overlapping submatrices/sliding window and vectorization problem

Hello experts, I have a colored (RGB) image which I want to do the following:

1. For each pixel I need to produce his 10X10 environment/neighbor pixels (lets say the pixel is in the middle) like sliding window. Thus, there will be overlaps of color images and that's what I want.
2. For each 10X10 environment (or color image) I need to produce a vector according to a specific logic I have.
3. Lastly, I need to save all vectors produced earlier for each pixel in some array/matrix (some structure).

FOR LOOPS are very slow in MATLAB. what you suggest me to do?

[lavanya p]

hello i need code for the following : 1)For each pixel i need to produce 10X10 environment/neighbor pixels (lets say the pixel is in the middle) like sliding window for a gray scale image . 2)or each 10X10 environment (or gray image) I need to produce glcm features of each pixel like correlation,energy,homogeneity,contrast. 3)Lastly, I need to save all vectors produced earlier for each pixel in some array/matrix (some structure)

[Image Analyst]

lavanya, I already gave an answer for this. Did you see it below? Did you run and adapt the demos? Use the one where I define a custom filter. In there you would put a call to graycomatrix().

[lavanya p]

thank u so much

Answer 1

[Image Analyst]

I think you haven't given the whole story,that's what I think. Who would want a gigantic list (perhaps 1 to 20 million) of 100 element vectors? What will they be then used for. To just give you the vectors, you can use blockproc(). I have demos. You can also use nlfilter() but that will probably be slower than a for loop.

AND can we dispel the myth that for loops are very slow? I can do a hundred million for loop iterations on my computer in 0.3 seconds. Would you call that slow? That's a lot of iterations in the blink of an eye. Sure, for some operations it can be slower than vectorized solutions but I'd hardly call them VERY SLOW - that's a myth. And most questions I see here do not have anywhere near that number of iterations, not even close.

[Roey Ginzberg]

Thanks for a quick answer!

I didn't understand how exactly I can retrieve all the sliding windows in one row of code (or a few...). I need a large list because I need to collect for every pixel- features that represents him and give this after to another algorithm- DBScan. Regards the slow time- I already tried it in FOR loops and it took seconds for every single row so that's why I posted this question. If something still not clear I will glad to clarify it. Again- thanks!

[Image Analyst]

See attached demos.