Integrating SIFT Features with Multivariate Gaussian Distribution to Enhance CBIR Performance
International Journal of Emerging Trends in Engineering Research · 2020
CBIR gained substantial importance in the recent fifteen years.Variety of modern systems have designed to effectively retrieve images from database and recognition of the objects in computer vision.Among all methods, Histogram of Gaussian (HOG) and Scale Invariant Feature Transform (SIFT) are standard procedures for attaining the features of the image.The image features achieved by SIFT are helpful in several applications, but these features prone to zero-order statistics as they derived from histogram quantities.This results in; there lacks the natural mechanism of high descriptiveness of image features.A novel method is developed based on Multivariate Gaussian distribution which establishes the relation among SIFT features in the neighbourhood.The exciting part of this approach lies in representing the non-linear Gaussian space into linear space as the Gaussian space is in Riemannian manifold.The Gaussian space can be mapped into linear Euclidean space using Lie group theory.The tests were conducted systematically on Caltech-101 and ZANG image databases to endorse this approach.The experiments demonstrated that a significant amount of the CBIR system performance has improvised with the proposed approach.