Abstract:
Computer vision has been an active research area for the past few decades. It is being used in many fields of industries for example digit recognition and handwriting recognition in banks, industrial parts recognition etc. Shape matching is an important sub-domain of Computer Vision, and relies on effective shape representation techniques. These techniques must be developed in order to effectively estimate similarity between the shapes. Major similarity estimation results correspond to effective shape matching techniques. In this thesis, we propose a multistage system using multiple feature representations for images. Every feature representation contributes its role towards matching and verifies and refines the results. Proposed method significantly speeds up complex but accurate shape matching techniques. Existing shape matching approaches concentrate on the accuracy perspective of shape matching without giving much consideration on the efficiency. Consequently, such approaches are accurate but cannot meet the performance demands online shape retrieval and classification. They are also not good for datasets containing extremely large number of shape samples. We handle this problem by presenting an extremely efficient shape matching approach. Selection of compressed, robust and accurate features is fundamental for effective content-based image recognition and retrieval using shape information of objects. In this paper, we present a three-stage system for object recognition and retrieval that employs multiple feature space representations of contour information. In the first stage, we pre-process the shapes to cater for the presence of distortions such as cracks that can significantly distort the contour information of the shape. We then generate multiple feature space representations of shapes to be used later in combination for efficient and accurate retrieval of shapes which includes compressed fourier coefficients. We then employ a given sophisticated but inefficient shape matching approach on the pruned dataset. We further combine our proposed Fourier descriptor based shape matching with a sophisticated shape matching approach to further enhance its accuracy. The proposed system is evaluated using publicly available shape datasets such as MPEG-7 and Swedish leaf datasets. Our approach achieves higher accuracies which are better than state-of-the-art approaches reported in literature with less retrieval time.
