Improved content-based image retrieval with multivariate Gaussian distribution

Abstract

This paper presents an overview of content-based image retrieval based on integrating image features obtained from scale-invariant feature transform (SIFT) with multivariate Gaussian distribution for the efficient retrieval of images in the image database. The state-of-the-art methods for image retrieval and object recognition use SIFT and HoG to extract image visual features. Though these descriptors are helpful in a variety of applications, they exploit zero-order statistics as they only collect histogram features, and this lacks high descriptiveness of object features and quantization problem. The novel method is having each pixel of the object which is associated with multivariate Gaussian distribution and approximated new features in the locality of the region. The key issue of this approach lies in space of the multivariate Gaussian distribution which lies in Riemannian manifold. But the linear space is suitable domain to discriminate image feature vectors efficiently. With the basis of Lie group structure and Riemannian geometry, multiplication operations are determined on the manifold to embed Gaussian space into linear space and are referred to as log-Euclidean multivariate Gaussian descriptors. These descriptors determine distinctive low-and high-dimensional image features efficiently. The experiments were conducted on Caltech-101, WANG database to validate thoroughly this approach.