Distance Transform

The example in this section is present in the source under mahotas/demos/distance.py.

We start with an image, a black&white image that is mostly black except for two white spots:

import numpy as np
import mahotas

f = np.ones((256,256), bool)
f[200:,240:] = False
f[128:144,32:48] = False

(Source code)

There is a simple distance() function which computes the distance map:

import mahotas
dmap = mahotas.distance(f)

Now dmap[y,x] contains the squared euclidean distance of the pixel (y,x) to the nearest black pixel in f. If f[y,x] == True, then dmap[y,x] == 0.

from __future__ import print_function

import pylab as p
import numpy as np
import mahotas

f = np.ones((256,256), bool)
f[200:,240:] = False
f[128:144,32:48] = False
# f is basically True with the exception of two islands: one in the lower-right
# corner, another, middle-left

dmap = mahotas.distance(f)

(Source code)

Distance Transform and Watershed

The distance transform is often combined with the watershed for segmentation. Here is an example (which is available with the source in the mahotas/demos/ directory as nuclear_distance_watershed.py).

(Source code)

The code is not very complex. Start by loading the image and preprocessing it with a Gaussian blur:

import mahotas
nuclear = mahotas.imread('mahotas/demos/data/nuclear.png')
nuclear = nuclear[:,:,0]
nuclear = mahotas.gaussian_filter(nuclear, 1.)
threshed  = (nuclear > nuclear.mean())

Now, we compute the distance transform:

distances = mahotas.stretch(mahotas.distance(threshed))

We find and label the regional maxima:

Bc = np.ones((9,9))
maxima = mahotas.morph.regmax(distances, Bc=Bc)
spots,n_spots = mahotas.label(maxima, Bc=Bc)

Finally, to obtain the image above, we invert the distance transform (because of the way that cwatershed is defined) and compute the watershed:

surface = (distances.max() - distances)
areas = mahotas.cwatershed(surface, spots)
areas *= threshed

We used a random colormap with a black background for the final image. This is achieved by:

import random
from matplotlib import colors as c
colors = map(cm.jet,range(0, 256, 4))
colors[0] = (0.,0.,0.,1.)
rmap = c.ListedColormap(colors)
imshow(areas, cmap=rmap)

API Documentation

A package for computer vision in Python.

Main Features

Compute global and local features (several submodules, include SURF and Haralick features)
Convolution and wavelets
Morphological features. Most are available at the mahotas level, include erode(), dilate()...
Seeded watershed implementation
read/write image


Coelho, Luis Pedro, 2013. Mahotas: Open source software for scriptable computer vision. Journal of Open Research Software, 1:e3, DOI: http://dx.doi.org/10.5334/jors.ac
mahotas.distance(bw, metric='euclidean2')

Computes the distance transform of image bw:

dmap[i,j] = min_{i', j'} { (i-i')**2 + (j-j')**2 | !bw[i', j'] }

That is, at each point, compute the distance to the background.

If there is no background, then a very high value will be returned in all pixels (this is a sort of infinity).

Parameters :

bw : 2d-ndarray

If boolean, False will denote the background and True the foreground. If not boolean, this will be interpreted as bw != 0 (this way you can use labeled images without any problems).

metric : str, optional

one of ‘euclidean2’ (default) or ‘euclidean’

Returns :

dmap : ndarray

distance map

Table Of Contents

Previous topic

Wavelet Transforms

Next topic

Polygon Utitilities

This Page