A Novel Non-Gaussian Feature Normalization Method and its Application in Content Based Image Retrieval

Abstract

In Content-Based Image Retrieval (CBIR) images are represented by multi low-level features that describe image color, texture, and shape of objects. The Efficient Manifold Ranking (EMR) algorithm is a semi-supervised learning algorithm on the low-level image features that has been used efficiently in CBIR. The combination of different image features to build the weighted EMR -graph usually uses normalized feature data for balancing the value of each feature. In this paper, we propose a novel normalization method for vector number data such as the low level image features where vector components are not consistent with the characteristics of the Gaussian distribution and its application for calculating the adjacent matrix of the weighted EMR-graph. Experiments show the effectiveness of the proposed algorithm for the EMR, the CBIR quality is really improved. Besides the testing normalization method for visual images, we also investigated the possibility to use the proposed method for medical image datasets.