Hi again, On Sun, Dec 16, 2012 at 7:09 PM, Parand Darugar <darugar@gmail.com> wrote:

Hello,

Not sure if it's appropriate to ask algorithm questions on this mailing list, please let me know if it's not.

Algorithm questions are definitely welcome.

I'd like to filter out large contiguous regions out of my image - I only care about smaller items. Following the Image Segmentation example/tutorial, I'm able to find and label the large segments using an Otsu threshold and ndimage.label.

Now I'd like to replace these large segments with "white", where "white" is the max value find in my grayscale image (basically I'd like to blank them out).

I'm not quite following how this is done in the tutorial:

...

...

...

label_objects, nb_labels = ndimage.label(fill_coins) sizes = np.bincount(label_objects.ravel()) mask_sizes = sizes > 20 mask_sizes[0] = 0 coins_cleaned = mask_sizes[label_objects]

How does the coins_cleaned = mask_sizes[label_objects] line clean the image?

In case you didn't follow the `bincount` part: The `sizes` array is a count of frequency of each label; so large labeled regions would have a large size. The index is matched to the integer label: the count for label `i` is stored in `sizes[i]`. `mask_sizes` just looks for all labeled region larger than 20 pixels and stores a boolean array (i.e. mask). So if the region labeled `i` has more than 20 pixels then `mask_sizes[i]` is True. Also, `i = 0` values are set to 0 because `i = 0` is reserved as the background label. Based on your question, you probably understood all that, but I just wanted to make sure we're on the same page. Now, the line that you actually asked about is a bit of a numpy trick: It uses `label_objects` to pick values from `mask_sizes`. A pixel labeled `i` (in `label_objects`) will grab the value in `mask_sizes[i]` (i.e. True or False); that value will then get placed in the output array at the same array position as the pixel in `label_objects`. That explanation was a bit confusing. Perhaps a short example will help:

...

...

import numpy as np values = np.array([0, 0, 0, 1]) # stand-in for `mask_sizes` picker = np.array([[0, 1, 2, 3], ... [3, 0, 1, 2], ... [2, 3, 0, 1], ... [1, 2, 3, 0]]) # stand-in for `label_objects` print values[picker] [[0 0 0 1] [1 0 0 0] [0 1 0 0] [0 0 1 0]]

All the values in the picker labeled 3 are set to 1 in the output because that was the value at `values[3]`. All other values were 0.

Generally speaking, given the labeled segments (label_objects from the tutorial), how would I apply operations to the original image - for example: set all of the pixels in "image" that are parts of objects labeled 3 and 9 to 0.577 .

Something like this should work:

...

...

mask_low = image < 3 mask_high = image > 9 image[mask_low & mask_high] = 0.577

Also, how do I get smarter about how the system works so I don't ask

elementary questions? Is there a good high level tutorial that'll familiarize me with the computation model for skimage / ndimage ?

Well, I think these questions are more focused on numpy, which forms the core of skimage. There are tons of numpy tutorials out there, but I don't have a particular favorite. Someone on the numpy mailing list pointed to the following tutorial: http://scipy-lectures.github.com/intro/numpy/index.html So that might be worth a look. Best, -Tony

Best,

Parand

--