Using the Time Warping Distance for Fourier-Based Shape Retrieval

Modern content-based image retrieval systems use different lowlevel features, like color, texture, and shape, to search in large image databases for those images which are perceptually similar to a given query image. Effective and efficient retrieval by shape similarity is still an open issue, despite the high importance of the shape feature in describing the content of an image. In this paper, we propose a new approach (based on the Discrete Fourier Transform) for assessing the shape similarity between two objects. The use of Fourier coefficients allows a compact representation of shape boundaries which is robust to noise and can be easily made independent of translation, scaling, rotation and changes in the starting point used to describe each boundary. Since the Euclidean distance, which is used by almost all Fourier-based approaches to shape-matching, is not effective whenever a phase shifting exists between the two boundaries to be compared, we propose the use of the (Dynamic) Time Warping distance to compare shape descriptors, allowing a (limited) elastic stretching of the time axis to accommodate possible phase differences. A comparative experimental analysis conducted on a real data set shows the superior effectiveness of our approach with respect to existing Fourier-based techniques.