An image is a matrix of colors. Each cell, called a pixel, contains a red, green, and blue value (mixing these can create any color). If we re-organize an image into a table such that we have one row per pixel and and red, green, and blue columns, then we can perform clustering on that table. The result will inform us about the main colors in the image.
This page has an image of the WI capital that we'll use: https://en.wikipedia.org/wiki/Madison,_Wisconsin. Download a resized version of it:
wget https://raw.githubusercontent.com/tylerharter/cs320/master/s21/lab13/capital.jpg
Run the following to read and view the image:
from matplotlib import pyplot as plt
img = plt.imread("capital.jpg")
plt.imshow(img)
img is three dimensional: row, column, color amounts. Check it's
.shape, then use .reshape fit it into a DataFrame like the one we
described above:
import pandas as pd
rgb = pd.DataFrame(????, columns=["red", "green", "blue"])
rgbANSWER
img.reshape(-1, 3)
Use KMeans clustering to identify 4 centroids for the colors:
from sklearn.cluster import KMeans
km = ????(n_clusters=4)
km.????(rgb)Look at km.cluster_centers_; this contains the four resulting 4
centroids (each on a row, each containing three color amounts).
Convert the centroid values to ints and place them in a DataFrame
called centroids that has the same columns as rgb.
ANSWER
pd.DataFrame(km.cluster_centers_.astype(int), columns=rgb.columns)
Reshape centroids.values so that we have 1 row by 4 columns by 3
colors, then pass the result to imshow:
plt.imshow(????)
ANSWER
centroids.values.reshape(1, 4, 3)
The result should look like this:
km = KMeans(n_clusters=4)
km.fit(img.reshape(????))
plt.imshow(km.cluster_centers_.reshape(????).astype(int))ANSWER
-1,3 and 1,4,3
How accurate is the image if we redraw it using only those 4 colors?
Complete the following using int, img.shape, labels_, and
cluster_centers_ (not necessarily in that order).
img2 = km.????.astype(????)[km.????].reshape(????)
plt.imshow(img2)