koenji.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-

import json
import urllib2
import urllib
import os
import glob
import PIL
import leargist
import numpy as np
import sklearn.decomposition
import sklearn.mixture

bing_id = YOUR_BING_ID
bing_key = YOUR_BING_KEY


def get_images(query):
    if not os.path.exists(query):
        os.mkdir(query)

    skips = [0, 50]
    count = 0
    for skip in skips:
        url = u'https://api.datamarket.azure.com/Bing/Search/Image?$format=json&Query=%27{{query}}%27&$skip={{skip}}'
        url = url.replace('{{query}}', urllib.quote(query.encode('utf-8', 'ignore')))
        url = url.replace('{{skip}}', str(skip))
        password_mgr = urllib2.HTTPPasswordMgrWithDefaultRealm()
        password_mgr.add_password(None, 'http://api.datamarket.azure.com', bing_id, bing_key)
        handler = urllib2.HTTPBasicAuthHandler(password_mgr)
        opener = urllib2.build_opener(handler)
        urllib2.install_opener(opener)
        line = urllib2.urlopen(url).read()
        open('%s/%d.txt' % (query, skip), 'w').write(line)
        r = json.loads(line)
        for item in r['d']['results']:
            img_url = item['MediaUrl']
            print count, img_url
            try:
                urllib.urlretrieve(img_url.encode('utf-8', 'ignore'), '%s/%d.jpg' % (query, count))
            except:
                print 'error'
                continue
            count += 1

def extract_features(directory):
    features = []
    filenames = glob.glob('%s/*.jpg' % directory)
    filenames.sort(lambda a,b: int(a[a.index('/')+1:a.rindex('.')])-int(b[b.index('/')+1:b.rindex('.')]))
    for filename in filenames:
        num = filename[filename.index('/')+1:filename.rindex('.')]
        try:
            im = PIL.Image.open(filename)
        except:
            print 'load error'
        features.append(leargist.color_gist(im))
    features = np.array(features)
    np.save('%s/features' % directory, features)

def kl(g1, g2):
    m1 = g1.means_[0]
    m2 = g2.means_[0]
    c1 = np.diag(g1.covars_[0])
    c2 = np.diag(g2.covars_[0])
    ret = np.log(np.linalg.det(c2)/np.linalg.det(c1))
    ret += np.trace(np.dot(np.linalg.inv(c2),c1))
    ret += np.dot(np.dot((m2-m1).T,np.linalg.inv(c2)), (m2-m1))
    ret += -len(m1)
    ret *= 1./2.
    return ret

def run():
    towns = [u'高円寺', u'下北沢', u'霞ヶ関', u'池袋', u'渋谷', u'浅草', u'新宿', u'神楽坂', u'代官山', u'表参道', u'原宿', u'バラナシ', u'デリー', u'ムンバイ', u'川越', u'京都']
    
    # get images
    for town in towns:
        get_images(town)
    
    # extract features
    for town in towns:
        extract_features(town)
    
    # load extracted features 
    d = []
    for town in towns:
        d.append(np.load(u'%s/features.npy' % town))

    # PCA
    pca = sklearn.decomposition.PCA(n_components=20)
    pca.fit(np.vstack(d))

    # GMM fitting
    g = []
    for i,town in enumerate(towns):
        g.append(sklearn.mixture.GMM(n_components=1, covariance_type='diag'))
        g[i].fit(pca.transform(d[i]))
    similarities = np.zeros((len(towns),len(towns)))

    # compute KL divergence
    for i in range(len(towns)):
        for j in range(len(towns)):
            similarities[i,j] = (kl(g[j],g[i])+kl(g[i],g[j])/2)

    # show results
    r = []
    for i in range(1,len(towns)):
        r.append((similarities[0, i], towns[i]))
    r.sort()
    for rr in r:
        print rr[1], rr[0]

if __name__ == '__main__':
    run()