A GENERIC RETRIEVAL METHOD USING PERCEPTUAL COLOR SPACE AND LOCAL DENSITY ESTIMATION H. Theoharatos 1 , S. Makrogiannis 1 , A. Ifantis 2 and S. Fotopoulos 1 1 : Electronics Laboratory, University of Patras, Patras 26500, Greece Tel: +3061 997465; fax: +3061 997456 2 : Technological Educational Institude of Patras, Greece {htheohar,macsoc,spiros}@physics.upatras.gr ABSTRACT In this work a color image retrieval scheme is proposed that combines color and shape statistical features for indexing. Concerning the color information, the perceptually uniform HSV color space has been employed. The shape feature is derived from the application of a non- parametric distribution modelling method on the original image. The weights of the color and shape fusion function were tuned using an objective evaluation method. The main advantages of this method are i) the modest computational complexity, ii) the small storage space requirements for feature indexing and iii) the efficient performance for various types of images. 1. INTRODUCTION During the last few years several image databases are being created due to the low cost of digital storage and the rapid growth of computational power. Automatic retrieval of an image from a whole database set is a relatively new research field. The limitations of conventional keyword- based methods, guided to the development of content- based retrieval systems (IR), such as the QBIC system by IBM [1]. These systems use low-level features such as color, shape and texture to represent the image content. Considerable research has been carried out on the basis of color and shape features. The comparison of images based on their histogram [2,3] is the most popular technique used for color retrieval. A lot of color spaces have been studied for retrieval purposes, that employ clustering [4] and segmentation [5] methods. These approaches although providing the overall system with better results in comparison to histogram-based techniques, are time consuming in the preprocessing stage. On the basis of shape retrieval, many recognition techniques have also been presented that use region- [6] or boundary-based methods [3,7]. The majority of the reported works deal with new methods for extracting color or shape information from an image. However, little attention has been given to the combination of low-level features [8,9], most of which has been carried out mainly on trademark and logotype image databases [3,10]. In our method, an integration of color and shape features was implemented to make a more flexible system and improve the retrieval performance, using a generic image database. A histogram technique was used on the perceptual HSV color space for color retrieval. The shape feature was extracted by means of an edge detection technique based on the estimation of the local density function (potential function), applied as a vector process [11,12]. 2. THE PROPOSED RETRIEVAL METHOD A typical retrieval system, in the preprocessing stage extracts the low-level features from the database images which are stored into indices. At the retrieval stage, an image usually queries the database and matching is calculated using a similarity function. The similarity measures are then ranked and the set of closest related images is presented as the output of the retrieval process. In order to build an efficient IR system, apart from performing well in the retrieval process we should take into consideration retrieval time, index-storage requirements and system’s flexibility. Histogram is used to describe the global pixel distribution of an image and its main advantage is attributed to its small sensitivity to variations in scale, rotation and translation of a given image. The use of histogram-based techniques yields in a simple retrieval model with little storage and computational time requirements. For the efficiency of our system, color and shape features are extracted from the database images using histogram techniques and the matching operation is performed using the histogram intersection (HI) operator [2] as a similarity measure, given by: