Clustering of Hand Written Digits Using K-Means Algorithm and Self Organizing Maps

Present work focuses on clustering of MNIST dataset using K-means clustering and Self-Organizing Maps (SOM). Histograms of Oriented Gradients (HOG) descriptors are used to extract the feature vectors and Principal Component Analysis (PCA) is applied on feature vectors to reduce the dimensionality. First two principal components are taken for cluster formation. Purity of cluster metric is used to evaluate the clusters. External criteria with prior information of true class is chosen to validate cluster. The performance of SOM is better than K-means in forming clusters. Out of 10 clusters K-means algorithm missed clusters of 3 digits (0, 7 and 9) whereas SOM missed clusters of 2 digits (5, 9).