Affine Invariant Approach for 2D Shape Matching (T-0662) (MFN 2936)

Abstract

The current ubiquity of digital image capturing systems has prompted a flurry of research activities aimed at the development of sophisticated content-based image data management techniques. The fundamental ingredient of content-based image retrieval is the selection of appropriate features to describe the content of the image. Shape of an object, represented by its contour, is the most important visual feature that is thought to be used by humans to determine the similarity of objects. The selected feature and its distance measure must be robust to different distortions such as noise, orientation, scale and rotation. In this thesis, we present an effective representation of shape, using its boundary information that is robust to arbitrary distortions and affine transformation. The contour representation of shape is converted into time series and then critical points are extracted from the time series using k-beam mechanism. The time series is partitioned between two consecutive critical points to facilitate partial matching. Partitions of time series are modeled using Discrete Fourier Transformation (DFT). Shape matching is then carried out in real space using Dynamic Time Warping. Angle between two consecutive critical points is calculated, compared during matching of shapes, and this leads to efficiency gains over existing approaches. Experimental evaluation demonstrates that the proposed shape representation and matching mechanism is effective, efficient and robust to different arbitrary and affine distortions.