Content-based Image Retrieval from Complete and Incomplete Sketches using Hierarchical Shape Descriptor

M.-W. Leung and K.-L. Chan (PRC)

Keywords

Image retrieval, incomplete user sketch, M-tree, Hierarchical Shape Descriptor

Abstract

Shape is one of the most basic and reliable features to describe objects. A Web-accessible content-based image retrieval by shape similarity is implemented in this project. Each image contains one or more objects. Object shape is segmented into tokens consisting of local features of maximum turn angle and orientation. User sketch is the query input. The retrieval algorithm matches the sketch with the objects in the database and generates a list of retrieved objects in the order of similarity to the user sketch. The similarity measure, called shape distance, is non-metric and resembles human perception. To enhance the system performance, the shape tokens are organised using M-tree indexing. Existing image retrieval systems by shape only rely on the outer boundary of object. It is obvious that the retrieval accuracy can be improved using both the outer boundary and the inner shape(s). A novel object feature, Hierarchical Shape Descriptor (HSD), is proposed. Our results show that HSD can easily distinguish objects with similar outer boundary but different inner shapes. Experiments are also carried out to evaluate the effectiveness of HSD in object retrieval under various degrees and in different layers of incomplete sketch. These situations can arise due to object occlusion or the flexibility of user sketch. Our results show that the system performs fairly accurate using various forms of incomplete query input.

Important Links:



Go Back


IASTED
Rotating Call For Paper Image