(IJCSIS) International Journal of Computer Science and Information Security,Vol. 08, No.05, 2010
Image Retrieval using Row Means of ColumnTransformed Even and Odd Parts of Image withWalsh, Haar and Kekre Transforms
, Sudeep D. Thepade
, Akshay Maloo
Ph.D.Research Scholar & Associate Professor,
B.Tech (CE) Student
Computer Engineering Department, MPSTME,
NMIMS (Deemed-to-be University), Mumbai, India
The paper presents an innovative content based imageretrieval (CBIR) technique based on row mean of columntransformed even and odd part of image. Here the performanceof proposed CBIR technique is tested for different transformslike Walsh transform, Haar transform and Kekre transform.Instead of using all pixels of image as feature vector for imageretrieval, row mean of transform applied on columns of odd andeven part of image is used, resulting into better performance withmuch lower computations. The proposed CBIR techniques aretested on generic image database having 1000 images spreadacross 11. For each transform 55 queries (5 per category) werefired. Then the precision and recall for all queries are computed.While considering the relevance of result images for calculationof precision and recall the results of odd image part and evenimage part both are ORED. The average precision and averagerecall values of all queries gives the performance comparison of proposed CBIR methods as compared to considering full imagedata as feature vector or considering the simple row mean asfeature vector. The results have shown the performanceimprovement (higher precision and recall values) with proposedmethods compared to all pixel data of image at reducedcomputations resulting in faster retrieval. The comparison of transforms for CBIR performance gives Walsh transformsurpassing the other two considered here.
Keywords-content based image retrival (CBIR), row mean,Walsh transform, Haar transform, Kekre transform.
The large numbers of images are being generated from avariety of sources (digital camera, video, scanner, the internetetc.) which have posed technical challenges to computersystems to store/transmit and index/manage image dataeffectively to make such large collections easily accessible.Image compression deals with the challenge of storage andtransmission, where significant advancements have been made[1,4]. The challenge to image indexing is studied in the contextof image database [2,6,7,10,11], which has become one of themost promising and important research area for researchersfrom a wide range of disciplines like computer vision, imageprocessing and database areas. The thirst of better and fasterimage retrieval techniques is increasing day by day. Some of important applications for CBIR technology could be identifiedas art galleries [12,14], museums, archaeology , architecturedesign [8,13], geographic information systems , trademark databases [21,22], weather forecast [5,21], medical imaging[5,18], criminal investigations , image search on theInternet [9,19,20].
Content Based Image Retrieval
In literature the term content based image retrieval (CBIR)has been used for the first time by Kato et.al. , to describehis experiments into automatic retrieval of images from adatabase by colour and shape feature. The typical CBIR systemperforms two major tasks [16,17]. The first being featureextraction (FE), where a set of features, called feature vector, isgenerated to accurately represent the content of each image inthe database. The second task is similarity measurement (SM),where distance between the query image and each image in thedatabase using their feature vectors is used to retrieve the top
“closest” images [16,17
]. For feature extraction in CBIR thereare mainly two approaches  feature extraction in spatialdomain and feature extraction in transform domain. The featureextraction in spatial domain includes CBIR techniques basedon histograms , BTC [1,2,16], VQ . The transformdomain methods are widely used in image compression, as theygive high energy compaction in transformed image [17,24]. Soit is obvious to use images in transformed domain for featureextraction in CBIR . But taking transform of image is timeconsuming, this complexity is reduced to a great extent by theproposed technique. Reducing the size of feature vector usingpure image pixel data in spatial domain only and till getting theimprovement in performance of image retrieval is the theme of the work presented. Many current CBIR systems use Euclideandistance [1-3,8-14] on the extracted feature set as a similaritymeasure. The Direct Euclidian Distance between image P andquery image Q can be given as equation 1, where Vpi and Vqiare the feature vectors of image P and Query image Qrespectively
with size „n‟.
The various transforms [5,24] used for proposed CBIRtechniques are discussed below: