Segmentation and Granulometry with scikit-image

[Tony Yu]

On Fri, Sep 21, 2012 at 8:46 PM, Stéfan van der Walt <stefan at sun.ac.za>wrote:

> On Wed, Sep 19, 2012 at 4:01 AM, ashz <ashz10001 at gmail.com> wrote:
>
> > Granulometry of Snowflakes:
> >
> http://www.mathworks.com/products/image/examples.html?file=/products/demos/shipping/images/ipexsnow.html
>
> Tony, do you think the morphological reconstruction would work for
> this?


Yes and no. Morphological reconstruction (specifically, h-dome, a.k.a.
h-maxima or regional max) tries to enhance features (maxima) based on the
local region. So the h-dome of a dim snowflake and bright snowflake will
look similar if they have a similar peakiness. In the original example, dim
peaks are still much dimmer than the bright peaks, even after the adaptive
histogram equalization.

Also, the h-dome of an image will tend to remove the corona-like blurring
around the larger snowflakes since the middle is so much brighter than the
surrounding region. I've attached an image of the results, along with the
code to generate it (although you can't run it without the snowflakes
image, which I probably shouldn't email). You'll notice that dim peaks are
much brighter than in the example, and the large snowflakes are smaller
(because of suppression of the corona region). These differences alter the
intensity curve quite a bit (I'm not sure what the "right" output is,
though).


> Otherwise perhaps one of the segmentation algorithms.
> Translating the example line-by-line is pretty easy, except that we do
> not have adaptive histogram equalization implemented yet.
>

BTW, there's a pending PR to add adaptive histogram
equalization<https://github.com/scikits-image/scikits-image/pull/242>.
(Although, I'm not sure how I feel about the large header file required for
that PR.)

As for the original post about these Matlab examples: Thanks for bringing
this up, ashz! These are exactly the type of examples I'd like to see in
the user guide: Something that takes the user all the way from image
loading to data extraction. We'd probably want to write our own instead of
simply translating the Matlab examples. It'd be great if people on this
list could provide examples from their own research.

Best,
-Tony


[image: Inline image 1]

# Code to generate image above
# Note: this won't run without image from Matlab install:
#     toolbox/images/imdemos/snowflakes.png


from skimage import io
from skimage import img_as_float
import skimage.morphology as morph
from skimage import exposure
from scipy import ndimage


# snowflake image from a Matlab install
img = io.imread('snowflakes.png')
# convert from uint8 to float since we'll be subtracting (prevent underflow)
orig = img_as_float(img)

img = exposure.equalize(img)
img = ndimage.gaussian_filter(img, 1)

# look for regional maxima greater than surrounding regions by h
h = 0.2
img_background = morph.reconstruction(img-h, img)
img_hdome = img - img_background
img = exposure.rescale_intensity(img_hdome)

io.imshow(img)
# io.imsave('snowflakes_enhanced.png', img)
io.show()
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scikit-image/attachments/20120924/1b8e9a73/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image.png
Type: image/png
Size: 16020 bytes
Desc: not available
URL: <http://mail.python.org/pipermail/scikit-image/attachments/20120924/1b8e9a73/attachment.png>