Computing, explaining and visualizing shape similarity in content-based image retrieval

Abstract

Although there is a growing need for Content-Based Image Retrieval systems, their use is often hampered by significant computational complexity and their inability to explain to their users the reasoning behind the similarity and retrieval processes they employ. This paper introduces Turning Function Difference (TFD), an efficient novel shape-matching method, which is based on the curvature of the shape outline and is translation, rotation and scale invariant. The method produces information about the correspondence of points belonging to the compared shapes that are used during the explanation process. TFD explains its results through an alignment and a visual animation process that highlights the similarities between the model images and each one of the selected images as perceived by the method. The proposed shape-matching method is used in the G Computer Vision (GCV) library, a single-object image retrieval system that utilizes information about the objects' outlines and explains the reasoning behind the selection of similar images to the user. The implemented system is freely available for download to all interested users.