B Ins a Pproach for Cbir by S Hifting the H Istogram to L Ower I Ntensities Using Proposed Polynomials

This paper describes the novel approach of feature extraction for CBIR systems. It also suggests the use of newly designed polynomial function to modify the image histogram so that the result of the CBIR system can be improved. To support this suggestion multiple polynomial functions have been tried. Out of which the best polynomial can be selected to modify the histogram for feature extraction. This gives better performance for the image retrieval based on contents. The separate histograms are obtained for each of the three color planes of the image so that the color information can be handled separately. These histograms are then divided into two equal parts by calculating the centre of gravity. This division of the R, G and B histograms into two parts lead towards the generation of eight bins. Eight bins are designed to hold different types of information like ‘Count of pixels’, ‘Total of intensities’ and ‘Average of intensities’. The work done includes the set of three polynomial functions used modify the histograms. Based on each polynomial function and the variation of the information used to represent the eight bin feature vector we could generate the multiple feature vector databases. Two types of bin sets based on type of the bin contents Total and Average of intensities with respect to each of the three polynomial functions for three colors we have 2x3x3 = 18 plus for count of pixels for each polynomial function giving 3 feature vector databases; like this total 18 + 3= 21 feature vector databases are prepared for the system testing. To demonstrate the performance of the system we have used database of 2000 BMP images from 20 different classes where each class has 100 images. 200 images are selected randomly as 10 images from each of the 20 classes to be given as query to the system. To compare the database and query image feature vectors and facilitate the retrieval three similarity measures are used namely Euclidean distance (ED), Absolute distance (AD) and Cosine correlation distance (CD). Performance of the system for all approaches discussed is evaluated using three parameters PRCP (Precision Recall Crossover Point), ‘Longest String’, and LSRR (Length of String to Retrieve all Relevant).