Classification of SURF image features by selected machine learning algorithms

We have proposed a concept for classification interesting points in images by means of a machine learning approach. The basic idea is that each interesting point detected in an image is classified either as a point belonging to some trained model (e.g. corner of a license plate) or not. During the first stage, we detected interesting points in a set of images by the well-known SURF method. Then we have employed supervised learning algorithms LDA, QDA, Naive Bayes, Decision tree and SVM to create relevant models of corners in images. Finally, all generated models were evaluated during classification stage by a cross-validation technique and an example experiment of license plate detection has been carried out and is introduced at the very end of this paper. Interesting outcomes have been obtained by the Naive Bayes algorithm resulting in a sensitivity value of the 100% and an accuracy value of the 99.8% on the real-world gallery of 535 images containing over 93 thousand interesting points. Although our gallery is not vast, the results are really promising to use our concept in another applications of robust and real-time object recognition.