Effect of Similarity Measures for CBIR Using Bins Approach

This paper elaborates on the selection of suitable similarity measure for content based imageretrieval. It contains the analysis done after the application of similarity measure namedMinkowski Distance from order first to fifth. It also explains the effective use of similarity measurenamed correlation distance in the form of angle ‘cosθ’ between two vectors. Feature vectordatabase prepared for this experimentation is based on extraction of first four moments into 27bins formed by partitioning the equalized histogram of R, G and B planes of image into threeparts. This generates the feature vector of dimension 27. Image database used in this workincludes 2000 BMP images from 20 different classes. Three feature vector databases of fourmoments namely Mean, Standard deviation, Skewness and Kurtosis are prepared for three colorintensities (R, G and B) separately. Then system enters in the second phase of comparing thequery image and database images which makes of set of similarity measures mentioned above.Results obtained using all distance measures are then evaluated using three parameters PRCP,LSRR and Longest String. Results obtained are then refined and narrowed by combining thethree different results of three different colors R, G and B using criterion 3. Analysis of theseresults with respect to similarity measures describes the effectiveness of lower orders ofMinkowski distance as compared to higher orders. Use of Correlation distance also proved itsbest for these CBIR results.