Evaluation of Different Metrics for Shape Based Image Retrieval Using a New Contour Points Descriptor

Abstract

In this paper, an image shape retrieval method was evaluated using Euclidean, Intersect, Hamming and Cityblock distances and different kinds of k-nearest neighbours classifiers such as the original kNN, mean distance kNN and Weighted kNN. Shapes were described using a new method based on the description of the contour points, CPDH36R, obtaining better results than with the original CPDH shape descriptor. The efficiency in the retrieval was tested using Kimia99, Kimia25, MPEG7 and MPEG2 datasets obtaining an 84% of success rate in Kimia25, 94% in Kimia99, 91% in MPEG2 and 82% in MPEG7 datasets using our CPDH36R method, cityblock distance and original kNN against the 68%, 91%, 74% and 59% respectively obtained using the original CPDH. The greatest difference between the original method and our proposal can be seen clearly using MPEG2 dataset. Another advantage of our retrieval method, apart from the success rate, is the computational cost which is clearly better than the one achieved with the original Earth Mover Distance classifier used in the CPDH original method.