State-of-Art: Similarity Assessment for Content Based Image Retrieval System

Abstract

CBIR (Content-Based Image Retrieval) is an approach in which similar images are retrieved for a particular query image depends on the image content similarity. The similarity measurement among the feature vectors and the feature mining process is the foremost image retrieval task. The performance of CBIR is depends on both ideal features mined from the image and the suitable choice of distance metrics (similarity measures). Various retrieval procedures have been established to improve the CBIR performances. This paper presents a summary of several similarity measures utilized in CBIR and the comparative study of these distance metrics on texture and color features. By HSV and color moments, color features are mined and texture features by Gabor wavelet filter (GWF). Similarity measures such as Manhattan distance (MHD), Cosine Angle distance (CAD), Euclidean distance (ED), Geodesic distance (GD), Minkowski distance (MKD), Jaccard distance (JD), Chebyshev distance (CD), Hamming distance (HD), Mahalanobis distance (MHBD), Earth mover distance (EMD), Kullback-Leibler divergence (KLD) and Bhattacharya distance (BD) were analyzed for feature similarity. The goal of this work is to analyze the performance of various similarity measures for CBIR. A systematic comparison of various similarity measures in the CBIR system for different texture and color feature vectors are done in this paper. The performance of two CBIR systems such as color and texture features (HSV, Moments, and GWF) and color and edge directive descriptor (CEDD) are computed. Using support vector regression with butterfly optimization algorithm (SVR-BOA) classifier, the performance of both the systems have been compared for different similarity features. The comparative results are tested on the WANG database and the performance of all these similarity measures in terms of precision, accuracy, and recall are evaluated for two CBIR systems.