Partitioning of Modified Histograms to Generate 27 Bins Feature Vector to Improve Performance of CBIR
H. B. Kekre1, Kavita Sonawane2
1Dr. H. B. Kekre, Computer Engg. Dept, NMIMS, MPSTME, SVKM, Mumbai, India.
2Ms. Kavita Sonawane, Computer Engg. Dept, NMIMS, MPSTME, SVKM, Mumbai, India.
Manuscript received on March 21, 2013. | Revised Manuscript received on April 15, 2013. | Manuscript published on April 30, 2013. | PP: 494-505 | Volume-2, Issue-4, April 2013. | Retrieval Number: D1491042413/2013©BEIESP
Open Access | Ethics and Policies | Cite
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Abstract: Content Based Image Retrieval is motivating the researchers to devise new techniques as the rate of retrieval is definitely gaining importance as multimedia databases are increasing day by day. In order to improve the retrieval accuracy of content-based image retrieval systems, research focus is on generating new efficient algorithms to extract image features and also to achieve the dimension reduction in order to reduce the processing time. In this paper, new algorithms are proposed to extract different types of image features based on the color contents of the RGB image. The feature extraction mainly deals with the original and the modified histograms of R, G and B planes. Four different modification functions namely Equalization (EQH), Logarithmic (LOG), Polynomial expression (POLY) and Linear equations 1, 2 and 3 (LINEQ 1, 2 and 3) are proposed to modify the histogram and their performance is compared in this paper. To implement the dimensionality reduction, this paper proposes two partitioning techniques namely Linear Partitioning (LP) and Centre of Gravity (CG)partitioning to partition the R, G and B histograms in three parts so that 27 bins can be generated from it. It directly reduces the size of the feature vector based on histogram from 256 bins to 27 bins only. Experimentation for the proposed methods is carried out using 2000 BMP images of 20 different categories. Comparison of query and database image feature vectors is worked out using three similarity measures namely Euclidean distance (ED), Absolute distance (AD) and Cosine Correlation distance (CD). To compare and evaluate the performances of all the proposed approaches along with different similarity measures three performance evaluation parameters are used namely Precision Recall Cross over Point , Longest String and Length of string to Retrieve all Relevant.
Keywords: About four key CBIR, Centre of Gravity, Equalization (EQH), Logarithmic (LOG), Polynomial expression(POLY), Linear Equations(LINEQ), Euclidean distance (ED),Linear Partitioning(LP), Centre of gravity(CG), Absolute distance (AD), Cosine Correlation Distance (CD), Precision Recall Cross over Point (PRCP), Longest String (LS), Length of String to retrieve all Relevant (LSRR).