$ CAT CASE-STUDY.MD · CLUSTERING · IMAGE COMPRESSION
Fashion MNIST clustering
and color-quantization compression.
Clustering techniques applied two ways: grouping similar clothing items in Fashion MNIST, and reducing the colors in an image to compress it.
01 · BREAKDOWN
Project breakdown.
Agglomerative clustering, Gaussian mixture models, and DBSCAN to explore patterns in Fashion MNIST, and K-Means color quantization to compress an image.
Explore clustering techniques
Group similar images and quantize colors with different methods.
Evaluate effectiveness
Assess how each method performs on Fashion MNIST.
Optimize compression
Reduce an image's colors while keeping visual quality.
02 · FINDINGS
Key findings.
GMM clusters looked most like real clothes.
Agglomerative clustering with 10 clusters gave the best separation. DBSCAN struggled (Rand Index 0.51–0.53). For compression, the elbow method picked K=4 as the best balance of size and fidelity.
Most interpretable clusters
Cluster centers resembled actual clothing items.
Inconsistent on images
Density-based clustering didn't suit this image data.
Efficient compression
K=4 was most efficient; K=10 preserved more detail.
