Hybrid classifier using SIFT descriptor

Abstract

In this paper, three image classification methods based on Interest Points (IP) will be presented to identify different types of noise attack (Affine, Zoom, Blur and Contrast) between two images using SIFT descriptor. The first one is an affine classifier deduced from several test on a standard images database. The second is a fuzzy classifier based on rules deduced from evolution curves of membership functions. Using the results obtained from the two classifiers, we conceive a third one based on their combination and denoted as "hybrid classifier". By analyzing the results of the three classifiers, we deduced that the last one leads to best results as we obtained a recognition rate exceeding 95%. This classifier will be considered as good solution to predict the noise attack type.