lnfmnatinn Pmcessing & Management, Vol. 33, No. 3, PP. 319-337. 1997 0 1997 Elsevier Science Ltd All rights reserved. Printed in Great Britain 0306-4573/97 $ I7 +O.OO zyxwvutsrqpon PII: SO306-4573(96)00069-6 zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQ SHAPE MEASURES FOR CONTENT BASED IMAGE RETRIEVAL: A COMPARISON BABU M. MEHTRE’, MOHAN S. KANKANHALLI’* and WING FOON LEE’ ’ New Technologies zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA Group, CMC Centre, Gachibowli, Hyderabad, 500019, India, ’ Institute of Systems Science, National University of Singapore, Kent Ridge, Singapore, 119597, Singapore and ) School of Electrical & Electronic Engineering, Nanyang Technological University, Nanyang Avenue, Singapore. 639798, Singapore (Received 2 October 1995; accepted 12 September 1996) Abstract-A great deal of work has been done on the evaluation of information retrieval systems for alphanumeric data. The same thing can not be said about the newly emerging multimedia and image database systems. One of the central concerns in these systems is the automatic characterization of image content and retrieval of images based on similarity of image content. In this paper, we discuss effectiveness of several shape measures for content based similarity retrieval of images. The different shape measures we have implemented include outline based features (chain code based string features, Fourier descriptors, UNL Fourier features), region based features (invariant moments, Zemike moments, pseudo- Zemike moments), and combined features (invariant moments & Fourier descriptors, invariant moments & UNL Fourier features). Given an image, all these shape feature measures (vectors) are computed automatically, and the feature vector can either be used for the retrieval purpose or can be stored in the database for future queries. We have tested all of the above shape features for image retrieval on a database of 500 trademark images. The average retrieval efficiency values computed over a set of fifteen representative queries for all the methods is presented. The output of a sample shape similarity query using all the features is also shown. 0 1997 Elsevier Science Ltd 1. INTRODUCTION There have been many studies in evaluation of information retrieval systems (Harman, 1992; Robertson & Hancock-Beauclieu, 1992; Salton, 1992) for alphanumeric data. Recently, there has been an increasing interest in the area of image and multimedia information systems. An important area of research in these emerging systems is the automatic characterization of image content and retrieval of images based on similarity of image content. In image information systems, a common criterion for data retrieval is “what objects (shapes or images) in the database match or are closely similar to a given shape or image ?’ This type of retrieval is called shape similarity based retrieval. Given the potentially large size of image databases resulting from the variety of visual sensors currently available, it seems reasonable to address the problem of shape matching with an eye to efficiency (Gary & Mehrotra, 1992). In this paper, we make an attempt to evaluate (assess the effectiveness) of several shape measures for similarity retrieval of images. Similarity retrieval based on image content forms an important feature for image and multimedia databases (Wu zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQ et al., 1995). In fact, there are a couple of commercial systems now available which have implemented variations of shape techniques described in this paper (Flickner et al., 1995; Bach et al., 1996). There are several shape measures (Marshall, 1989; Mehtre & Narasimhalu, 1992) which can be used for recognition and retrieval of images, an overview of which is given in Section 3. An important criterion for shape measures is that the measures be invariant to affine transformations (i.e. rotation, scaling, and translation) of images. This is because, human beings ignore such * To whom all correspondence should be addressed. IP” 13:3-c 319