Research On The Feature-point’s Detectors Of Digital Images

With rapid development of computer technology and artificial intelligence, the extracted image feature information has increasingly been important in the field of digital image processing. In many of the characteristics of image, feature point which represents the local structure has played a unique role. It is a key step to extract the feature points in many applications such as image matching, motion tracking, object recognition, machine navigation and so on. So the feature-point’s detectors have been attented by many experts and many excellent detectors have been come out.In this thesis, we first analyze the classic feature-point’s detectors, point out the strengths and weaknesses of the various algorithms. And we find that the weaknesses of Harris method and Harris-Laplace method are come from the Gauss Filter which has weak the details of some weak feature points, and make these points can’t been responsed by the operators. Besides, Harris-Laplace operator remove some weak feature points which been detected in Harris method under low scales and cannot remove some redundant feature points which been strong responsed by using a unified LoG operator.In the process of feature point detection, image filtering operation plays a very important role. Based on the research of the filtering operation, we found that the bilateral filter preserve the edge feature of the images, which fits well with the feature point detection. However, the bilateral filter is not easy to be realzed by hardware. Therefor, the Thiele’s continued fraction is used to approximate exponential function, which effectively improves the running speed.Finally, this thesis presents an improved feature-point’s detection. We check the feature points with multi-scale Harris method by using the inproved bilateral filter to replace the Gauss filter, then track and group it while checking feature points so that some points that represent the same local structure are divided into a group, then select a point to represent the structure which has the closest eigenvalues of the auto-correlation matrix from each group. It is shown by experiments that the method can get good effect in the feature point recall, precision and the remove redundant feature points.