Υλοποίηση υπερ-ταξινομητών πολυμεσικών δεδομένων
Κάτσαρης, Δημήτριος Α.
This graduate dissertation is dealing with the case of classification of patterns which belong to more than two classes. The design of the super-classifier was based in a method called "one against all". According to this method, for each data class, there is created a classifier that is trained by using positive and negative training data population. Positive data consist of snapshots of the class, which the classifier corresponds with each time, while negative data consist of snapshots of remaining classes. If the population of positive data and the population of negative data is equal, then we have balanced classifier, otherwise, we have an unbalanced classifier. After the integrated training of all the particular classifiers, we can conclude which class a give snapshot possibly belongs to by checking this snapshot with all the above particular classifiers. Based on this method, we created our own models that assisted us to design the final model of our super-classifier. Initially, the approach of the problem was based on the use of the classifier SVM (Support Vector Machines). However, in order to have a safer conclusion for the method "one against all", we also used neural networks that unfortunately returned worse results during the evaluation process. The evaluation of the above design methods of the super-classifier was focused on three particular classification problems. The two of them refer to classification problems of musical pieces, for which suitable characteristic vectors have been computed, and the third one refers to the classification of human face expressions. The results show that the final hybrid design method of the super-classifier seems to have slightly better performance that of the pure (original) method "one against all". Also, it seems that the quality of the set of characteristics, which are given as an input to the classifier, play a considerable role in the deduction of the final results.