The Magic of Principal Component Analysis through Image Compression

Utilizing Images to Beautifully Demonstrate PCA

The Magic of Principal Component Analysis through Image Compression

What is PCA?

Principal Component Analysis or PCA is a dimensionality reduction technique for data sets with many continuous (numeric) features or dimensions. It uses linear algebra to determine the most important features of a dataset. After these features have been identified, you can use only these features to train a machine learning model and improve performance without sacrificing accuracy. As a good friend and mentor of mine said:

“PCA is the workhorse in your machine learning toolbox.”

PCA finds the axis with the maximum variance and projects the points onto this axis. PCA uses a concept from Linear Algebra known as Eigenvectors and Eigenvalues. There is a post on Stack Exchange that beautifully explains it.

Image Compression

PCA is nicely demonstrated when it’s used to compress images. Images are nothing more than a grid of pixels and a color value. Let’s load an image into an array and see its shape. We’ll use imread from matplotlib.

import numpy as np
import matplotlib.pyplot as plt
from matplotlib.image import imread

image_raw = imread("cat.jpg")

(3120, 4160, 3)


Image by Author

The results show a matrix of size (3120, 4160, 3). The first is the height of the image, the second is the width, and the third is the three channels of RGB values. Given the number of dimensions of this image, you can see how compared to a classic tabular set of data would be considered quite large, especially when we think about the 3,120 columns.

Before we continue, let’s change this to a grayscale image to remove the RGB value.

# Show the new shape of the image
image_sum = image_raw.sum(axis=2)

# Show the max value at any point. 1.0 = Black, 0.0 = White
image_bw = image_sum/image_sum.max()

(3120, 4160)

Calculating Explained Variance

Next, we can fit our grayscale image with PCA from Scikit-Learn. After the image is fit, we have the method pca.explained_variance_ratio_, which returns the percentage of variance explained by each principal component. Utilizing np.cumsum we can add up each variance per component until it reaches 100%. We'll plot this on a line and show where 95% of explained variance would be.

import numpy as np
from sklearn.decomposition import PCA, IncrementalPCA

pca = PCA()

# Getting the cumulative variance
var_cumu = np.cumsum(pca.explained_variance_ratio_)*100

# How many PCs explain 95% of the variance?
k = np.argmax(var_cumu>95)
print("Number of components explaining 95% variance: "+ str(k))

plt.title('Cumulative Explained Variance explained by component')
plt.ylabel('Cumulative Explained variance (%)')
plt.xlabel('Principal components')
plt.axvline(x=k, color="k", linestyle="--")
plt.axhline(y=95, color="r", linestyle="--")
ax = plt.plot(var_cumu)
Number of components explaining 95% variance: 54

Image by Author

First, I want to point something out. By printing off the length of components, we can see that there are 3120 components overall, showing how the number of components relates to the width of our image.



And by plotting this, we can see how dramatically the curve accelerates towards 100% and then flattens. What’s crazy is that we only need to use 54 of the original 3120 components to explain 95% of the variance in the image! That's quite incredible.

Reducing Dimensionality with PCA

We’ll use the fit_transform method from the IncrementalPCA module to first find the 54 Principal Components and transform and represent the data in those 54 new components. Next, we’ll reconstruct the original matrix from these 54 components using the inverse_transform method. And finally, we’ll then plot the image to assess its quality visually.

ipca = IncrementalPCA(n_components=k)
image_recon = ipca.inverse_transform(ipca.fit_transform(image_bw))

# Plotting the reconstructed image
plt.imshow(image_recon,cmap =

Image by Author

We clearly can see that the quality of the image has been reduced, but we can identify it as the original image. When PCA is applied along with Machine Learning models such as image classification, both training times are reduced dramatically, and prediction times on new data produce nearly as good results but with fewer data.

Showing other Values for k-Dimensions

Next, let’s iterate over six different k-values for our image, showing the progressively improving image quality at each number. We’ll only go to 250 components, still just a fraction of the original image.

def plot_at_k(k):
    ipca = IncrementalPCA(n_components=k)
    image_recon = ipca.inverse_transform(ipca.fit_transform(image_bw))
    plt.imshow(image_recon,cmap =

    ks = [10, 25, 50, 100, 150, 250]


    for i in range(6):
    plt.title("Components: "+str(ks[i]))

    plt.subplots_adjust(wspace=0.2, hspace=0.0)

Image by Author


And that’s it! As few as 10 components even let us make out what the image is, and at 250 it's hard to tell the difference between the original image and the PCA-reduced image.

PCA is an extremely powerful tool that can be integrated into your workflow (via pipelines) to dramatically reduce the number of dimensions in your dataset without much loss of information. Keep in mind that PCA is designed for continuous or numeric data use. Check out this article, PCA Clearly Explained, for more details and the math behind PCA. Thanks for reading, and enjoy!