Affine Invariant Feature Extraction And Recognition Of Shapes

Shape recognition is an attracted area in the field of computer vision and image understanding. After reviewing the shape analysis technique and analysing the difficulties in shape recognition caused by changes of view point, changes of illumination, noise, and occlusion, this dissertation focuses on the technique of affine invariant feature extraction and feature-based shape matching, which is based on the the physic model of affine transform between shapes.Similar invariant feature is an important part of affine invariant feature, which can be categorized into contour-based feature and the region-based feature. For contour-based feature extraction, traditional methods are commonly sensitive to noise. To solve this problem, An even-grid-polar mapping is applied and the idea of statistic is introduced, then the invariant extremum feature is proposed based on the the distribution of geometric features of a shape. For region-based feature extraction, traditional methods commonly need shape normalization to achieve similar invariances and thus will induce some errors to shape features. To overcome this drawback, a multilevel similar invariant shape feature is proposed, which utilizes the properties of Radon transform and exhibits sufficient descriptive capability by characterizing a shape in different levels inherently.Compared with the similar invariant features, entirely global affine invariant features are more applicable. Since an affine transform group is a Lie group, affine invariant feature extraxtion can be investigated based on group theory. Traditional methods based on Lie group theory use derivatives as coordinates to prolong the group action thus the resulting affine invariant features will depend on derivatives and are sensitive to noise. To solve the problem, a novel parameterized shape description named integral bending function (IBF) is presented, which is derived by extending a Lie group action on R 2 to potential jet space. IBF can be calculated by integrating operation and thus robust against noise. An efficient algorithm for computation of IBF is given and the properties of IBF are analyzed and proved in theory. For an affine parameterized shape, IBF is an affine invariant feature after normalizing by shape area. For an arc-length parameterized shape, the local peaks of IBF correspond to shape critical points, and thus IBF can the used for shape critical points extraction.For non-occluded shape matching, the problem is how to effectively measure the similarity of the extracted affine invariant features. According to the influence of shape noise and digitalization on affine invariant features, a notion of distance between shapes based on feature regulating and attention mechanism is defined. The proposed distance function can be implemented by dynamic programming, and its simple form for measuring the distance of invariant extremum feature is given. Compared with shape matching methods based on the classical similarity measures, the proposed method has advantanges in robustness and applicability.For occluded shape matching, traditional decomposition-based methods have difficulties in obtaining the stable local parts of a shape. Aiming at this problem, a new occluded shape matching method is proposed, which is based on stable local parts of a shape defined by identical shape decomposition. Firstly, an affine invariant descriptor named area ration sequence (ARS) is constructed based on breakpoints to describe the spatial relations of a shape, which reduces the problem of identical decomposition to the problem of finding the longest common substring of ARS. Secondly, the multiple local elastic matching algorithm is designed, which can find stable longest common substring of ARS effectively. Finally, through affine invariant description for stable local parts defined by shape identical decomposition, affine invariant local feature is achieved and then for matching shapes. The proposed method is able to deal with inconsistent shape decomposition, occlusion, affine deformation and robust to nosie.The researches will provide important theoretical basis for the designment and implementation of a shape-based atuomatical target recognition system.