An improvement of dissimilarity-based classifications using SIFT algorithm

Abstract

In dissimilarity-based classifications (DBCs), classifiers are not based on the feature measurements of individual objects, but rather on a suitable dissimilarity measure among the objects. In this paper, we study a new way of measuring the dissimilarity between two object images using a SIFT (Scale Invariant Feature Transformation) algorithm [5], which transforms image data into scale-invariant coordinates relative to local features based on the statistics of gray values in scale-space. With this method, we find an optimal or nearly optimal matching among differing images in scaling and rotation, which leads us to obtain dissimilarity representation after matching them. Our experimental results, obtained with well-known benchmark databases, demonstrate that the proposed mechanism works well and, compared with the previous approaches, achieves further improved results in terms of classification accuracy. © 2011 Springer-Verlag Berlin Heidelberg.