Research On Image Invariant Feature Detection And Description

Local invariant feature extraction plays an important role in the research of computer vision, and it has been applied to many fields successfully, such as image registration, target recognition, texture recognition, scene recognition and so on. In practical applications with different kinds of images, local invariant feature methods need to be of low complexity, low storage, high repeatability and high distinctiveness. In this thesis, for the sake of fulfilling the demand, some fundamentals and key technologies related to principle of local invariant feature extraction, have been systematically studied on feature detection, the creation of nonlinear scale space and feature description.Section of local invariant feature detection,gives an overview of basic concepts, and summarizes the performance of different algorithms of detection. We may encounter some difficulties as the selection criterion of Harris corner may ignore some real corners. In order to overcome the problem, a selection threshold for Harris corners is proposed. The threshold selection is based on the motivating factors of stability and repeatability and can be optimized by using the ratio of two eigenvalues. And the means proposed can be extended to many other selection criterions based on the determinant and trace of the matrix. When constructing scale space, a nonlinear scale space is built using efficient Additive Operator Splitting technology and variable conductance diffusion. This nonlinear scale space due to an relative high distinctiveness performs much better than linear ones,.Further research is done in the feature detection approaches by using the oriented histogram information, to bridges the gap between local feature detectors and descriptors. The algorithm avoids the step of description, and exhibits high performance in terms of repeatability, distinctiveness and low computation complexity.Section of the feature description, many state-of-the-art descriptors including SIFT, Principal Component Analysis SIFT, GLOH, DAISY and the machine learning descriptors, have been analyzed. A new framework has been constructed to improve the capability of DAISY. Firstly, Principal Component Analysis is applied to reduce the diminutions of feature description vector. Then the vector is normalized, which permits the modified method to exhibit high performance in low storage.