applying an image filtering technique on 'selected' pixels

Hi,

The feature extraction step of my algorithm, applies some filter on data and use the value of filtered pixels as the feature of original pixels of image.

My problem is that I need the feature of only small subset of pixels(threshold-ed pixels), not all of the image. and it is time consuming to filter all of the image, instead of only some pixels.

My question is how can I filter only selected pixels of an image? Is there any matlab function for this purpose?(I think I can not use imfilter)

Thank you.

Answer 1

[Image Analyst]

How time consuming is it? Give me a time in seconds. Cropping out all the small regions to their own subimages and then filtering, and then sending them back is one option but it may not be any faster.

[leila]

using tic & toc function, the filtering time is "0.335012 seconds", number of thresholded pixels is at least 300.

[leila]

The image is 3D

[Walter Roberson]

What order of elapsed time are you hoping to get it down to? How many hours or days of your time will you spend to reduce the elapsed time by 1/10th of a second?

[Image Analyst]

Personally I wouldn't worry about it. It's fast enough and not worth the trouble, unless you want to do video rate processing on a live image. Does the 3D mean color? Or is it a volumetric gray scale image (like CT or MRI)?

[leila]

it is volumetric gray scale image: 3D US image.

[leila]

yes, I must reduce the run-time. it must be used in real-time applications.

[leila]

So, after separating subimages and filtering, the result may not faster than imfilter...

[leila]

although I don't expect real-time functionality from the matlab codes, I try to write efficient code.

[Image Analyst]

Well, you can try what I said but I think with all the figuring out where contiguous blocks live, then extracting them, then filtering them, and then assigning them back to their original locations, and all the bookkeeping involved with that, it won't be more efficient. I think it will be a lot more complicated, harder to maintain, and in the end, probably not faster. But you're welcome to try. If you know that you're always going to filter one known constant block, like image3d(20:50, 42:69, 123:173), then it might not be too bad, but if you had dozens of irregularly shaped and connected blobs, then it could be complicated. And that's not even considering what you're going to do with the edge effects (boundary effects) where your filter kernel partially covers the voxels that you want to filter and partially includes voxels that you don't want to filter.

[leila]

Thank you for excellent guidance, your comments help me a lot.