Non negative matrix factorisation in python on individual images

Question

I am trying to apply NMF to a particular image that is loaded in grayscale mode. I have tried several links but my image after application of NMF remains almost the same and cannot be distinguished with the grayscale image initially loaded.

However, when i come across the scikit-learn's code on implementing decomposition on a dataset, i see that the faces there have been transformed into ghost - like faces. Here is the link:

http://scikit-learn.org/stable/auto_examples/decomposition/plot_faces_decomposition.html#sphx-glr-auto-examples-decomposition-plot-faces-decomposition-py

And here is the code I am using:

import cv2    
from sklearn import decomposition    
import matplotlib.pyplot as plt    

img = cv2.imread('test1.jpeg',0)    
estimator = decomposition.NMF(n_components = 2, init = 'nndsvda', tol = 5e-3)    
estimator.fit(img)    
vmax = max(img.max(), -img.min())    
plt.imshow(img, cmap=plt.cm.gray, interpolation = 'nearest',vmin=-vmax,vmax=vmax)    
plt.show()

I am new to the techniques of NMF on matrices espicially such a large image numpy array.
My image is test1.jpeg that is 225 * 224 .jpeg image.

Can someone please help me on implementing the code for a single image? Thanks a lot in advance.

Answer 1

The reason you get the original image in the plot is that you actually plot the original image. Instead you would need to work with the output of estimator.

The NMF decomposition produces two matrices W and H that compose the original matrix. You need to multiply those to get the image.

import cv2    
from sklearn import decomposition    
import matplotlib.pyplot as plt 
import numpy as np   

img = cv2.imread('data/trump2.jpg',0)  
vmax = max(img.max(), -img.min())

fig, (ax, ax2)  =plt.subplots(ncols=2)    
ax.imshow(img, cmap=plt.cm.gray, interpolation = 'nearest',vmin=-vmax,vmax=vmax)

n_components = 20

estimator = decomposition.NMF(n_components = n_components, init = 'random', tol=5e-3)    
W = estimator.fit_transform(img)
H = estimator.components_

new_img = np.dot(W,H)
ax2.imshow(new_img, cmap=plt.cm.gray,
                   interpolation='nearest',
                   vmin=-vmax, vmax=vmax)

plt.show()