Mahalanobis-Constrained Hard-Clustering Algorithm

The paper proposes a modification of the standard k-means algorithm that removes low-probability tail points of Gaussian components using the p-quantile of the chi-squared distribution as a trimming threshold based on the squared Mahalanobis distance. The procedure is evaluated on synthetic noisy datasets and compared with well-known state-of-the-art clustering methods. Applications in circular image pattern recognition demonstrate effective detection of spherical patterns in highly noisy environments. Pattern detection is driven by an optimization using a constructed spherical clustering-validity index, and the method’s practical potential is confirmed by comparisons with state-of-the-art circular detection techniques.