Research On Feature Extraction And Matching Of Images

Image matching is a hot and difficult research topic in the field of digitalimage processing. With the development of computer technology andmatching theory, the range of applications of image matching is getting widerand wider, gradually developed from the original military field to human dailylives and industrial production. Matching technology has very importantsignificance, which constitutes the foundation of the technologies such asimage stitching, image recognition, image retrieval, moving target tracking,etc..This dissertation systematically studies matching techniques based onimage features with focus on the algorithms of the image feature pointdetection and methods of the image feature descriptor.First，this dissertation describes the basic principles and preprocessing ofimage matching algorithm, and then scale space theory. The traditionalGaussian image pyramid is established by convolution of the original image.We improve the traditional method by establishing the next layer of image byconvolution of the anterior layer, which reduces computations.In terms of feature point detection, a variety of feature points extractingalgorithm is introduced. There are redundant feature points when multi-scaleHarris method is adopted to check. To address this issue, this dissertation hasproposed an improved method of checking Harris-Laplace feature points, thattrack and group the Harris-Laplace feature points so that some feature pointsthat represent the same local structure are divided into a group, then use theGaussian differential function to remove redundant feature points and thecorner measure to select the most representative of the local structure of thecorner.In terms of descriptors of the feature point, this dissertation analyzes theSIFT matching method in details. Inspired by the Harris operator and the SIFTdescriptor method, we use the correlation matrix of the Harris region todescribe each region characteristics from vectors of personal change and mutual change, and use zoning policies of SIFT descriptor to build a point descriptor.This descriptor is a116-dimensional vector, which will not be affected byimage rotation and brightness.