Intento de k-means para segmentación de imágenes

kmeans.py

'''
Created on Mar 17, 2013

@author: JuanTejada
'''
from math import sqrt
from operator import div
import Image
import random
import sys
MAX_ITER = 20000
MIN_DIFF = 1

class Point():
    
    def __init__(self, coords = (0,0,0), count = 1, dim = 3, matcoords = (0,0)):
        self.coords = coords
        self.count = count
        self.dim = dim
        self.matcoords = matcoords
    
    def __eq__(self, other):
        return self.coords == other.coords
    
    def __str__(self):
        return str(self.coords)
    
    def __repr__(self):
        return str(self.coords)
    
    def __hash__(self):
        return hash(self.coords)

def main():
    path = "../../input/imagenes/42.png"
    k = 10
    #path = input("Enter the absolute path for the image:")
    #k = input("Enter k:")
    img = Image.open(path)
    #img.show()
    img.thumbnail((200, 200))
    mat = img.load()
    points = getPoints(img.size, mat)
    clusters = kmeans(points, k)
    print "Final Centroids: "+ str(clusters.keys())
    
    #colors = [(0,0,0),(255,255,255),(255,0,0),(0,255,0),(0,0,255),(127,127,127)]
    i = 0
    for cluster in clusters:
        for point in clusters[cluster]:
            mat[point.matcoords] = tuple(map(int,cluster))
        i += 1
    img.show()
    
def kmeans(points, k):
    centroids = generateSeeds(k, points)
    print "Seeds: "+ str(centroids)
    
    old = []
    clusters = {cent.coords:[]  for cent in centroids}
    ptdim = getPointDim(points)
    diff = 0
    count = 0
    while True:
        print "k: "+str(len(centroids))
        
        for point in points:
            clusters[minDistanceCentroid(point, centroids)].append(point)
        old = centroids
        centroids = calculateCentroids(clusters, ptdim)
        diff = max(diff, max([distance(old[i].coords, centroids[i].coords) for i in range(len(centroids))]))
        
        if sameCentroids(old,centroids) or diff < MIN_DIFF or count>MAX_ITER:
            break
        else:
            clusters = {cent.coords:[]  for cent in centroids}
            count += 1
    
    return clusters

def distance(p1,p2):
    return sqrt(sum([((p1[i]-p2[i])**2) for i in range(len(p1))]))

def calculateCentroids(clusters, ptdim):
    centroids = []
    for cluster in clusters.values():
        ptct = 0
        acum = [0 for i in range(ptdim)]
        for point in cluster:
            ptct += point.count
            for i in range(len(point.coords)):
                acum[i] += point.coords[i]
        centroids.append(Point(tuple(map(div, acum, [float(ptct)]*ptdim)), 1, ptdim))
    return centroids

def minDistanceCentroid(point, centroids):
    min = sys.maxint
    cen = 0
    for i in range(len(centroids)):
        d = distance(point.coords, centroids[i].coords)
        if d < min:
            min = d
            cen = i
    return centroids[cen].coords

def generateSeeds(k, points):
    s = set(random.sample(points, k))
    while len(s)!=k:
        s = set(random.sample(points, k))
    return list(s)

def getPointDim(points):
    return points[0].dim

def getPoints(size, imgMat):
    points = []
    x, y = size
    #counts = {}
    for i in range(x):
        for j in range(y):
            points.append(Point(imgMat[i,j], 1, len(imgMat[i,j]), (i,j)))
    #for count, coords in img.getcolors(x * y):  
        #points.append(Point(coords, count, len(coords)))
    return points

def sameCentroids(old, new):
    for o in old:
        if not o in new: return False
    return True
    
main()