Fusion in phase space for shape retrieval

2-D shape recognition and content-based image retrieval using the boundary information as a 1-D time series is a common way of addressing the shape-matching problem. A significant factor for successful object identification is insensitivity to basic shape deformations, such as scaling, shifting, rotation, partial distortions etc. In this paper, a method for handling the time series boundary information as a set of feature vectors in the reconstructed phase space is proposed. Besides enhanced discrimination, projection in phase space facilitates the fusion of several boundary descriptors as well. Optimal information extraction from the multivariate time series is achieved by proper embedding dimension and time-lag parameter selection. Centroid contour distance (CCD) and Angle Sequence (AS), capturing both local and global shape information in a circular manner, are the descriptors employed in the fusion stage. Invariance to rotations and insensitivity to partial deformations is achieved while retrieval performance is enhanced. The method is applied and evaluated in MPEG-7 part B shape database using early or late fusion.