Query-by-sketch based image retrieval using diffusion tensor fields

Abstract

A user-drawn sketch is one of the most intuitive forms of Human Computer Interaction. Users can express their intention by sketching the specific characteristics of a target object as a rough and simple black and white hand-drawn draft image. Recent advances of tablet PC and multi-touch screen technology raised increasing interest on how users might search and retrieve the desired images in databases from a simple sketched image. In this paper, we present a new approach for content based image retrieval from a query by sketchy draft images which are not in the database. Our innovation to sketch based image retrieval systems consists of three steps: (i) Image database configuration using size normalization, edge detection, and hierarchical image classification, (ii) Tensorial feature extraction of query and image data in the topology of second-order symmetric diffusion tensor fields, and (iii) Similarity measure using eigen-features between sketched query and databases to retrieve the most similar target object. Experiments are conducted to evaluate the performance of our methodology showing an efficient and mature image retrieval system.