Covariance based Clustering for Digital Image Compression

Abstract

Traditionally clustering is done based on distance between the data points and the cluster centroids. A new type of clustering based on distance between the covariance matrices of the clustered data points was recently introduced. This allows clustering data points according to their spatial distribution model. This paper presents an improved algorithm for the computation of covariance based clustering with application to digital image compression. Karhunen-Loeve Transform is an optimal decorrelating transform. But its application in image compression is limited due to the fact that it is dependent of the covariance of the image data and the computations are expensive. The proposed method clusters image blocks according to induced covariance matrices and applies pre-computed Karhunen-Loeve transform basis matrices. Experimental results on CT and MRI images show the superior compression and image quality measures.