Statistical approach to clustering in pattern recognition

Clustering is a typical method of grouping data points in an unsupervised learning environment. The performance of most clustering algorithms is dependent on the accurate estimate of the cluster number, which is always unknown in the real applications. We propose a parametric approach, which starts with an estimate of the local distribution and efficiently avoids pre-assuming the cluster number. This clustering program is applied to both artificial and benchmark data classification and its performance is proven better than the well-known k-means algorithm.