Effective image database search via dimensionality reduction

Abstract

Image search using the bag-of-words image representation is investigated further in this paper. This approach has shown promising results for large scale image collections making it relevant for Internet applications. The steps involved in the bag-of-words approach are feature extraction, vocabulary building, and searching with a query image. It is important to keep the computational cost low through all steps. In this paper we focus on the efficiency of the technique. To do that we substantially reduce the dimensionality of the features by the use of PCA and addition of color. Building of the visual vocabulary is typically done using k-means. We investigate a clustering algorithm based on the leader follower principle (LF-clustering), in which the number of clusters is not fixed. The adaptive nature of LF-clustering is shown to improve the quality of the visual vocabulary using this. In the query step, features from the query image are assigned to the visual vocabulary. The dimensionality reduction enables us to do exact feature labeling using kD-tree, instead of approximate approaches normally used. Despite the dimensionality reduction to between 6 and 15 dimensions we obtain improved results compared to the traditional bag-of-words approach based on 128 dimensional SIFT feature and k-means clustering.