This is used to illustrate the process of this classification processing.
In this experiment, I did it in the following steps.
1.I downloaded the images from Zoo Galaxy by the coordinatoors of galaxies. The data source is in file 'GalaxyZoo1_DR_table7.csv' and the code is in file 'downloadGalaxy.py'(Please create an account in http://www.sciserver.org/ and replace variables 'Authentication_loginName' and 'Authentication_loginPassword' with your own username and password).
2.I used the code in repo https://github.com/CarolineHaigh/mtobjects to detect the galaxies size(or say the pixel positions) in each of the images.
So in this step you don't need to run the lines of code in mto.py
mto.generate_image(image, id_map, params)
mto.generate_parameters(image, id_map, sig_ancs, params)
By using the id_map generated in code
id_map = mto.relabel_segments(id_map, shuffle_labels=False)
you could find out the id that contains the galaxy.
The follow is what I did with the id_map: I know it's not a good solution (even a bit of verbosing). So please optimize it if you would like to do(which is also should be done).
object_ids = id_map.ravel()
sorted_ids = object_ids.argsort()
id_set = list(set(object_ids))
# # print ("the length is %d" % len(id_set))
right_indices = np.searchsorted(object_ids, id_set, side='right', sorter=sorted_ids)
left_indices = np.searchsorted(object_ids, id_set, side='left', sorter=sorted_ids)
assumeXstart = 0
assumeXend = 0
assumeYstart = 0
assumeYend = 0
assumeArea = 0
assumeYs = []
assumeXs = []
for n in range(len(id_set)):
pixel_indices = np.unravel_index(sorted_ids[left_indices[n]:right_indices[n]], processed_image.shape)
width = np.amax(pixel_indices[1]) - np.amin(pixel_indices[1])
height = np.amax(pixel_indices[0]) - np.amin(pixel_indices[0])
if (width >= 180) and (height >= 180) and (width != 1023) and (height != 1023):
currentCenterX = (assumeXstart + assumeXend) / 2.0
currentCenterY = (assumeYstart + assumeYend) / 2.0
newCenterX = (np.amax(pixel_indices[1]) + np.amin(pixel_indices[1])) / 2.0
newCenterY = (np.amax(pixel_indices[0]) + np.amin(pixel_indices[0])) / 2.0
print ("the widthh is %d and height is %d" % (width, height))
if (pow(abs (currentCenterX - 512) , 2) + pow(abs(currentCenterY - 512) , 2)) > (pow(abs (newCenterX - 512), 2) + pow(abs(newCenterY - 512) ,2)) :
assumeXstart = np.amin(pixel_indices[1])
assumeXend = np.amax(pixel_indices[1])
assumeYstart = np.amin(pixel_indices[0])
assumeYend = np.amax(pixel_indices[0])
assumeArea = len(pixel_indices[0])
assumeYs = pixel_indices[0]
assumeXs = pixel_indices[1]
print ("The n is %d" % n)
3.Then I just cut only the areas contain the galaxies from each images.(I used the variables from last step - including assumeXstart, assumeXend, assumeYstart, assumeYend, assumeYs, assumeXs).
4.I built max-trees based on the new generated images(contains only the galaxies). In the meantime, I calculated all the moments in each of node in max-trees and generated the local spectrums.
Again, I used the code from Caroline Haigh. But I replaced the original maxtree.h and maxtree.c to my new C files 'MTMaxTree.h' and 'MTMaxTree.c'. In these two files, I calculated the moments of each node during the constructing of max-tree and save the local spetrums to local files.
5.By the local spectrums generated from step 4, I used the GMLVQ(http://matlabserver.cs.rug.nl/gmlvqweb/web/) to train the model and observe the results.
For now I just classified two classes - elliptic and spiral. I think it would be better if SBa, SBb,SBc.. or Sa, Sb, Sc.. could be classified by using this approach.