Histogram Bins Matching Approach for CBIR Based on Linear grouping for Dimensionality Reduction

This paper describes the histogram bins matching approach for CBIR. Histogram bins are reduced from 256 to 32 and 16 by linear grouping and effect of this dimensionality reduction is analyzed, compared, and evaluated. Work presented in this paper contributes in all three main phases of CBIR that are feature extraction, similarity matching and performance evaluation. Feature extraction explores the idea of histogram bins matching for three colors R, G and B. Histogram bin contents are used to represent the feature vector in three forms. First form of feature is count of pixels, and then other forms are obtained by computing the total and mean of intensities for the pixels falling in each of the histogram bins. Initially the size of the feature vector is 256 components as histogram with the all 256 bins. Further the size of the feature vector is reduced to 32 bins and then 16 bins by simple linear grouping of the bins. Feature extraction processes for each size and type of the feature vector is executed over the database of 2000 BMP images having 20 different classes. It prepares the feature vector databases as preprocessing part of this work. Similarity matching between query and database image feature vectors is carried out by means of first five orders of Minkowski distance and also with the cosine correlation distance. Same set of 200 query images are executed for all types of feature vector and for all similarity measures. Performance of all aspects addressed in this paper are evaluated using three parameters PRCP (Precision Recall Cross over Point), LS (longest string), LSRR (Length of String to Retrieve all Relevant images).