Content based image retrieval using multiple features and Pareto approach

Abstract

In this paper, we propose a method for image retrieval based on the Pareto method. Each Pareto point is represented by a vector of distances between image features. In the method, feature distance can be applied with the use of any existing distance measures such as euclidean distance, Histogram Intersection distance, etc. A point is called a Pareto point if and only if there does not exist any other point that is less than or equal for all dimensions of distance measure with respect to the input image. The set of all Pareto points forms a set of fronts at different depths. We also propose formal properties of the Pareto front. Specifically, we prove the Pareto front depth with regards to a point in the search space is the number of vertices of the longest dominant path.
A content based image retrieval (CBIR) system executes an image classifier according to the input query image. The classification engine takes the set of all fronts as the original dataset set. Relevance feedback is used to integrate user’s information feedback over returned results. The experimental result on three image collections shows the effectiveness of the proposed method by reducing noise data for supervised machine learning classification engines or avoiding local strap in query refinement techniques such as query point movement and query expansion. Our algorithm called PDFA (Pareto Front Depth Algorithm) uses a flexible threshold to get Pareto points of k-depths and it saves search space up to 70\%.