LalithShiyam / unsupervised-image-clustering

An unsupervised image clustering algorithm that uses VGGNet for image transformation. Python, scikit-learn and tensorflow.

Unsupervised Image Clustering using ConvNets and KMeans algorithms

This is my capstone project for Udacity's Machine Learing Engineer Nanodegree.

• For a full description of the project proposal, please see proposal.pdf.

• For a full report and discussion of the project and its results, please see Report.pdf.

• Project code is in capstone.ipynb

Brief Description

I theorised that we can use KMeans clustering to seperate unlabelled images of different entitites after using ConvNets to transform them into a more meaningful representation.

I use the convolutional layers of Keras's VGGNet model with ImageNet weights to transform cat and dog images. I then use Principal Component Analysis (PCA) for dimensionality reduction, before passing the new representation to a KMeans clustering algorithm for seperation (labelling).

This is implemented in capstone.ipynb where you can find more details about what the code is achieving.

Results have been successful with up to 97.7% accuracy achieved.

Libraries you will need

All code is in Python 3.6.

• cv2
• keras
• matplotlib
• numpy
• pandas
• sklearn
• tensorflow