Fashion MNIST & Lena Image Clustering

Exploring Clustering Techniques for Image Analysis

This project investigates clustering techniques for image analysis using Fashion MNIST and Lena images.

Project Breakdown

This project investigates the use of clustering techniques in two different applications:

1. Clustering for Data Exploration - Using Fashion MNIST, we group similar clothing items to understand patterns in the dataset.

2. Image Compression via Color Quantization - Using the Lena image, we apply clustering to reduce the number of colors, achieving compression.

We employ Agglomerative Clustering, Gaussian Mixture Models (GMM), and DBSCAN for image clustering, and K-Means for image compression.

Investigate clustering methods for grouping images and color quantization.

Assess how different clustering methods perform on Fashion MNIST and Lena images.

Reduce the number of colors in Lena image while maintaining visual quality.

Agglomerative Clustering with 10 clusters provided the best separation in Fashion MNIST.
Gaussian Mixture Models (GMM) captured meaningful cluster centers, with improved interpretability.
DBSCAN struggled with Fashion MNIST, producing moderate Rand Index scores (0.51 - 0.53).
K-Means compression (K=4 & K=10) demonstrated a trade-off between visual quality and compression.
The Elbow Method identified K=4 as optimal for balancing compression and image fidelity.

GMM produced the best-defined clusters for Fashion MNIST, resembling actual clothing items.

DBSCAN performed inconsistently on image data, with moderate Rand Index scores (~0.52).

K-Means compression efficiently reduced colors, with K=4 resulting in high efficiency and K=10 preserving more detail.