Classifiers combination for recognition score improvement
Résumé
In this report we describe two widely used combinational methods: stacked generalization and respectively the behavior knowledge space in order to improve the recognition scores obtained with the multi-layer perceptron and respectively the support vector machines. By the combination of these two classifiers we achieved good results, taking in consideration the drawbacks, in the database were low represented classes, and in the mean time the quality of the images was not sufficiently good. By introducing some regroupment operations on the classes respectively some rejection criterias on the final decision rules, we raised satisfactory recognition scores. We found that the stacked generalization method is giving better results than the BKS method which can be explained with the complexity of the final decision rules used in these two methods. As future works, we suggested to introduce some new classifiers in order to improve the combinational process in the BKS method and to refine the normalization process for the real images by addign more local information in the trandformation process.
Loading...