International Journal of Computer Applications (0975 – 8887) Volume 99– No.6, August 2014 48 Comparative Evaluation of Transform and Cluster based CBIR Seema Anand Chaurasia Assistant Professor Computer Department Xavier Institute of Engineering, Mahim, Mumbai, India Omprakash Yadav Assistant Professor Computer Department Xavier Institute of Engineering, Mahim. Mumbai, India Vaishali Suryawanshi Assistant Professor Computer Department Thadomal Shahani Engineering College, Bandra, Mumbai, India ABSTRACT In this paper, comparative evaluation is made on the result of CBIR system based on transform based Image retrieval and cluster based Image retrieval by taking Euclidean distance as similarity measure to calculate the deviation from query image. In transform based image retrieval discrete wavelet transform is used to decompose the image. The image is decomposed till sixth level and last level approximate component is saved as feature vector. For cluster based Image retrieval LBG algorithm is used. Comparisons are made between results of transform based Image retrieval and cluster based Image Retrieval. In this paper two experiments are carried out on COIL database images, first on 720 images having 10 different classes and second on 1440 images of 20 different classes. General Terms Average Precision, Average Recall, Query Image, Database, Similarity Measurement. Keywords CBIR, QBIC, Precision, Recall, Query Image, Precision, Recall, HAAR, LBG, Euclidean Distance. 1. INTRODUCTION An image retrieval system is effective if it can retrieve an image or a collection of images from an operating database upon being inputted with a single image whose replicas or lookalikes need to be extracted. Database management and Computer vision are two major research communities, which study the subject of image retrieval from different perspectives. Text based image retrieval employs techniques of attaching text or data along with the image to describe it, often termed as ‘metadata’, while content based techniques use visual features to match images to the query image. In CBIR, all the images from the database are taken and features of each are extracted and stored in a vector. These features are compared to the extracted features of the query image. A CBIR typically converts images in feature vector representations and uses them to match similar images [1]. IBM was the first research company to take initiative by proposing a system called QBIC (query by image content), which was developed, at the IBM Almaden Research Center. Unlike keyword based system, visual features are extracted from images itself. Content based image retrieval system uses contents and extracts features like color, shape and texture. All these are visual contents [2]. Feature extraction based on the color is done taking color histogram of each image. It is nothing but the portion of pixels within an image which has some specific value. This specific value people express as colors. One more benefit of using extraction based on colors is that it does not depend on the size of image. Eventually, color histograms will be taken and compared [3].Feature extraction based on shape does not refer to the shape of the whole image. Within an image we have certain area of interest. Shape denotes the shape that area of interest. To get the shape first image segmentation is performed or edge detection is done [3]. Feature extraction based on texture measurably concentrates on visual patterns and how they are organized. It is given by texels. Texels are kept into number of different sets, depending how many of textures detected. The sets determined like this will give complete information about which patterns and where in the image they got detected. The determination of specific texture can be made by molding texture as a 2D gray level variation. For determination of level of contrast, regularity, directionality and coarseness pixels relative brightness is considered. In this paper texture features are extracted using DWT transform in Transform based retrieval [3]. Figure shows a general description of a standard image retrieval system. Fig 1: Basic block diagram of CBIR system As shown in the above diagram, feature extraction is done for both query image as well as database images. First of all, all the databases images will be presented and there feature will be extracted and stored as feature database. Then query image is selected and its features will be extracted. This process will result into query features. After this process, query feature is taken and all the database feature is selected, and depending on criteria similarity measure will be calculated and the retrieved images will be displayed. In this paper, similarity measurement criteria is taken based on Euclidean distance [3].