Neat problem!

For feature extraction, `skimage.feature` is probably your friend. Nothing against integral images, but I'm not sure they are going to give you an ideal feature set for discrimination (you can see that visually). Also, attempting to normalize your input data might be worth looking into at some point as it appears exposure is not uniform.

As a first pass, you could feed raw grayscale values straight into e.g. a Bernoulli Restricted Boltzmann machine, or check out scikit-learn's excellent tutorial on digit recognition. Though for both of those options, the performance is going to be strongly dependent on the quality and - especially - quantity of the training set.

Beyond that, thresholding and a skeletonization with `skimage.morphology.skeletonize` might give you informative morphology data to feed in to a classifier. 

Best of luck,
Josh

On Tuesday, February 24, 2015 at 10:21:10 AM UTC-6, Jean-Patrick Pommier wrote:

Dear All,

I am trying to make pairs of images from the following set of images (chromosomes sorted by size after rotation). The idea is to make a feature vector for unsupervised classification (kmeans with 19 clusters)


From each chromosome an integral image was calculated:

plt.figure(figsize = (15,15))
gs1 = gridspec.GridSpec(6,8)
gs1.update(wspace=0.0, hspace=0.0) # set the spacing between axes.
for i in range(38):
   # i = i + 1 # grid spec indexes from 0
    ax1 = plt.subplot(gs1[i])
    plt.axis('off')
    ax1.set_xticklabels([])
    ax1.set_yticklabels([])
    ax1.set_aspect('equal')
    image = sk.transform.integral_image(reallysorted[i][:,:,2])
    imshow(image , interpolation='nearest')

Then each integral image was flatten and combined with the others:

Features =[]

for i in range(38):
    Feat = np.ndarray.flatten(sk.transform.integral_image(reallysorted[i][:,:,2]))
    Features.append(Feat)
X = np.asarray(Features)
print X.shape

The X array contains 38 lines and 9718 features, which is not good. However, I trried to submit these raw features to kmeans classification with sklearn using a direct example :

from sklearn.neighbors import NearestNeighbors
nbrs = NearestNeighbors(n_neighbors=19, algorithm='ball_tree').fit(X)
distances, indices = nbrs.kneighbors(X)
connection = nbrs.kneighbors_graph(X).toarray()
Ploting the connection graph shows that a chromosomes is similar to more than one ...
  • Do you think that integral images can be used to discriminate the chromosomes pairs?
  • If so, how to reduce the number of features to 10~20? (to get a better discrimination)
Thanks for your advices.

Jean-Patrick