I wanted to see if I could use ML algorithms to create vibrant pop art. The goal was simple: get an image, find clusters for its pixels, and recolour.

K-Means Clustering

K-means algorithm groups datasets into clusters, where k is the number of clusters. Below are some data points clustered into 4 groups.

Our future pop art

I generated this cartoon image of a boy at a bus stop. For simplicity’s sake, we’ll be using a greyscale version for our calculations.

For the greyscale image, each pixel is represented by a value from 0 to 255 with 0 being pitch black and 255 being white. We can therefore use these values as our data points to cluster with K-Means.

Below is a histogram of the frequency of each pixel intensity in the image from pitch black to pure white.

How many clusters?

No matter how many clusters we choose, our solution isn’t going to be perfect. With fewer clusters, we lose more information but have more stark grouping. As we increase our clusters, each cluster becomes less potentially less meaningful but we see more nuance. If we had the full 256 clusters, one for each pixel intensity, then the clusters become meaningless.

To showcase this, the following images show our greyscale image clustered into 2 to 9 clusters respectively.

To optimise our clusters, we’ll use both inertia and silhouette scores.

Inertia Score measures how condensed and tightly packed the clusters are. The lower the score, the more tightly packed they are. However, as we have more clusters this number decrease naturally, so low isn’t simply better. We’d like a distinct point where adding more clusters doesn’t give us much more gains in inertia.

Silhouette Score shows how strongly data points belong to their cluster and distinctly not to another. If many data points sit between two or more clusters groups and could nearly belong to another, then our clusters don’t capture our data well. As such, a higher silhouette score is better.

The inertia score’s rate of improvement tends to slow around k=4. Simultaneously our silhouette score drops from 4 to 5. As such, let’s move ahead with 4 clusters.

Clustering and Colouring

Recall our pixel intensity distribution from before? Here it is again, followed by the data points split into 4 clusters and then recoloured with some pop art colours.