A new approach for unsupervised classification

Abstract.

Pattern recognition generally requires the description of objects in terms of a set of measurable features. The quality of features representing each object has a considerable bearing on the success of pattern classification. Generally, the feature selection and clustering analysis are processed separately and most clustering techniques consider all features with equal importance. This paper presents a new approach of unsupervised classification, which examines each feature separately in accordance to a specific order predefined by the algorithm. This approach is mainly based on mode detection of unidimensional histograms corresponding to each feature. The results obtained on real and simulated test data are presented.