In [1]:

    

import io
import urllib
import collections
import logging
import warnings

import numpy as np
import matplotlib.image
import matplotlib.pyplot as plt
import matplotlib.gridspec

%matplotlib inline

import pandas
import skimage
import skimage.util
import skimage.data
import skimage.color

import skimage.segmentation
import skimage.feature
import sklearn.cluster

warnings.filterwarnings('ignore')


    

In [2]:

    

# use a painting
url = 'https://upload.wikimedia.org/wikipedia/commons/7/7b/Vincent_van_Gogh_-_The_yellow_house_%28%27The_street%27%29.jpg'
# specify how many colors we want
n_colors = 15


    

In [3]:

    

# read the data
f = io.BytesIO(urllib.urlopen(url).read())


    

In [4]:

    

# convert in memory to an image
img = matplotlib.image.imread(f, format='jpeg')


    

In [5]:

    

# downscale to at ...x1000
img = skimage.transform.rescale(img, 1000.0/max(img.shape))
plt.imshow(img)
plt.tight_layout()


    

In [6]:

    

# about 20 segments in x and y
n_segments = 20*20 # separate image into n segments


    

In [7]:

    

labels = skimage.segmentation.slic(img, n_segments=n_segments, compactness=20, enforce_connectivity=True) # split up
shuffled = np.arange(n_segments) # create labels
np.random.shuffle(shuffled) # and shuffle them
colors = matplotlib.cm.Accent( matplotlib.colors.Normalize(vmax=n_segments)(shuffled)) # lookup a color


    

In [8]:

    

# this does not mix in the background
plt.imshow(colors[labels]) 
# this does
plt.imshow(skimage.color.label2rgb(labels, img, colors=colors, alpha=0.5))


    

    
Out[8]:





<matplotlib.image.AxesImage at 0x104196e50>






    

In [9]:

    

regionprops =  skimage.measure.regionprops(labels+1)
slices = [np.s_[props.bbox[0]:props.bbox[2], props.bbox[1]:props.bbox[3]] for props in regionprops]


    

In [10]:

    

rows = []
for i, (region, slice) in enumerate(zip(regionprops, slices)):
    row = dict(label= i, color = np.mean(img[slice][region.image], axis=0), x=region.centroid[0], y=region.centroid[1])
    rows.append(row)


    

In [11]:

    

df = pandas.DataFrame.from_records(rows)
mean_colors = np.array(list(df['color'].apply(list)))


    

In [12]:

    

fig, ax= plt.subplots(figsize=(20,12))
ax.imshow(mean_colors[labels])
# or we could have just done
# ax.imshow(skimage.color.label2rgb(labels, img, kind='avg'))


    

    
Out[12]:





<matplotlib.image.AxesImage at 0x11f97ca10>






    

In [13]:

    

kmeans = sklearn.cluster.KMeans(n_clusters=n_colors)
kmeans.fit(mean_colors)
colorgroup = kmeans.predict(mean_colors)
colorgroups = kmeans.cluster_centers_


    

In [14]:

    

plt.subplots(figsize=(20,12))
plt.imshow(colorgroups[colorgroup[labels]])
for i, row in df.iterrows():
    plt.text(row['y'], row['x'], colorgroup[i])


    

In [15]:

    

rgb = colorgroups[np.newaxis,:,:]
fig, ax = plt.subplots(figsize=(20,3))
ax.imshow(rgb, interpolation='none')


    

    
Out[15]:





<matplotlib.image.AxesImage at 0x10cd42f50>






    

In [16]:

    

hsv = matplotlib.colors.rgb_to_hsv(rgb)
sortidx = np.argsort(hsv.squeeze()[:,0], axis=0)


    

In [17]:

    

fig, ax = plt.subplots(figsize=(20,3))

ax.imshow(rgb[:,sortidx,:], interpolation='none')
_ = ax.set_xticks(np.arange(sortidx.shape[0]))
_ = ax.set_xticklabels(sortidx)


    

In [18]:

    

counter = collections.Counter(colorgroup)
pandas.DataFrame(data=dict(color=counter.keys(), n=counter.values())).sort('color')


    

    
Out[18]:






  

    

      

      
color
      
n
    
  
  

    

      
0
      
0
      
3
    
    

      
1
      
1
      
31
    
    

      
2
      
2
      
30
    
    

      
3
      
3
      
12
    
    

      
4
      
4
      
28
    
    

      
5
      
5
      
12
    
    

      
6
      
6
      
45
    
    

      
7
      
7
      
11
    
    

      
8
      
8
      
87
    
    

      
9
      
9
      
6
    
    

      
10
      
10
      
5
    
    

      
11
      
11
      
35
    
    

      
12
      
12
      
36
    
    

      
13
      
13
      
11
    
    

      
14
      
14
      
31
    
  

In [53]:

    

color_idx = np.argsort(sortidx)

fig, axes = plt.subplots(2, 1, figsize=(13,13), gridspec_kw=dict(height_ratios=[4,1]))

ax = axes[0]
ax.set_xticks([])
ax.set_yticks([])
ax.imshow(skimage.segmentation.boundaries.find_boundaries(labels), cmap='Greys')
for i, row in df.iterrows():
    ax.text(row['y'], row['x'], color_idx[colorgroup[i]] + 1)

ax = axes[1]qs
ax.imshow(rgb[:,sortidx,:], interpolation='none')
ax.set_xticks([])
ax.set_yticks([])
for i, idx in enumerate(color_idx[sortidx]):
    ax.text(i, 0, str(idx + 1))
fig.tight_layout()


    

In [43]:

    

def closest_color(rgb):
    """find the closest matching rgb label"""
    lab1 = skimage.color.rgb2lab(rgb[np.newaxis, np.newaxis, :])
    labels = [label 
              for label in dir(skimage.color.rgb_colors) 
              if not label.startswith('_')]
    rgbs = [getattr(skimage.color.rgb_colors, label) 
            for label 
            in labels]
    labs = [skimage.color.rgb2lab(np.array(rgb, dtype='float32')[np.newaxis, np.newaxis, :]) 
            for rgb
            in rgbs]
    distances = [skimage.color.deltaE_ciede94(lab1, lab2)
                 for lab2 in labs]
    idx = np.argmin(distances)
    return labels[idx]


    

In [ ]: