Similarity measures for Content-Based Dermoscopic Image Retrieval: A comparative study

Similarity measures play crucial role in Content-Based Dermoscopic Image Retrieval (CBDIR). This paper analyses and compares images based respectively on twelve distances namely: Minkowski, Euclidean, Standardized Euclidean, Mahalanobis, Manhattan, Chebychev, Cosine, Canberra, Relative Deviation, Bray-Curtis, Square Chord and Square Chi-Squared measures for CBDIR. Two dermatologists were asked to diagnose 176 skin lesion images in order to classify them. Eight common classes of pigmented skin lesions have been identified, including: Melanoma, Nevus/Mole (ML), Lentigo (Len), Basal Cell Carcinoma (BCC), Seborrhoeic Keratosis (SK), Actinic Keratosis (AK), Angioma (AG) and Dermatofibroma (DF). Color and texture features have been extracted from the segmented skin lesions. Then a series of CBDIR experiments were conducted on the image database. The results indicate that the CBDIR performance is significantly improved by using Canberra and Bray-Curtis distances compared to conventional measures.

I. Introduction

Currently, we are witnessing the era of image and multimedia revolution, thanks to the tremendous progress in technologies that allow easy and affordable image capturing and sharing. The web is actually such a technology that contributed to a large extent in the increase of this phenomenon. As a result, huge image databases have been produced over the last decades, and with this new environment, effective and efficient image browsing as well as searching tools are needed.