Simple and fast shape based image retrieval

Abstract

Content Based Image Retrieval (CBIR) is a challenging and active topic of research. This paper focuses on the shape of the represented objects as the main criterion in respect to evaluate the relevance of the retrieval results. There are several shape descriptors described in the literature, which are reported to achieve good results. However some of them are not obvious to implement or are computationally demanding. In this paper we propose a simple and fast to compute set of features to achieve shape based image retrieval. We conducted retrieval experiments on the MPEG-7 Core Experiment CE-Shape-1 test set and the results obtained demonstrate usefulness and competiveness against reported results from other more elaborated descriptors. Results demonstrate that our approach is also valuable when objects represented in the images share similar shapes, although being conceptually different. Another interesting result is that users tend to be very stringent when a good result set is presented (very similar shapes) whilst they are more permissive when the result set does not present a very high level of similarity between the shapes.