Trademark retrieval by relaxation matching on fluency function approximated image contours

Abstract

We propose an automatic retrieval method that addresses the problem of finding similar trademark images from a database when compared with an input. Our method is based on evaluating the compatibility of relative relations among extracted image contour segments between the input and the registered images using relaxation matching. The overall compatibility, after finding a best matching configuration between contour segments of the input and each registered image, provides the basis for a distance metric used in similarity ranking. Experiments were carried out on 300 randomly selected trademark images obtained from the Japan Patent Office's Website for performance evaluation. The applicability of our method, however, is not limited to trademark retrieval but can be extended to other applications having the need for image similarity matching.