CLASSIFICATION OF DIFFERENT TYPES OF CORONAS USING PARAMETRIZATION OF IMAGES AND MACHINE LEARNING
Igor Kononenko, Matjaž Bevk, Sasa Sadikov, Luka Šajn · 2004
We describe the development of computer classifiers for various types of coronas. In particular, we were interested to develop an algorithm for detection of coronas of people in altered states of consciousness (two-classes problem). Such coronas are known to have rings (double coronas), special branch-like structure of streamers and/or curious spots. Besides detecting altered states of consciousness we were interested also to classify various types of coronas (six-classes problem). We used several approaches to parametrization of images: statistical approach, principal component analysis, association rules and GDV software approach extended with several additional parameters. For the development of the classifiers we used various machine learning algorithms: learning of decision trees, naive Bayesian classifier, K-nearest neighbors classifier, Support vector machine, neural networks, and Kernel Density classifier. We compared results of computer algorithms with the human expert’s accuracy (about 77% for the two-classes problem and about 60% for the six-classes problem). Results show that computer algorithms can achieve the same or even better accuracy than that of human experts (best results were up to 85% for the two-classes problem and up to 65% for the six-classes problem).