Research On Corner Detection Algorithms Based On Image Contour

Corner is one of the main features of objects on the image contour. Corner detection is a basic research topic in the area of computer vision and image processing. Precise corner detection is very important in many computer vision tasks such as image matching, target recognition and motion estimation. Images in real life may be interfered by noise and background, or may be disturbed by visual angle, light, scale, translation, rotation and affine. Corner detection helps to remove the fore mentioned interferences and disturbances. The efficiency and performance of image processing is highly dependent on the corner detector. It is a critical problem of how to improve the precision and speed of corner detection algorithm to recognize corners of objects and keep robustness.This paper is funded by Chongqing Nature and Science Fund grant"Reasearch on Image Processing technology basic application"which aims to propose various feature detector. So the main work in this paper is to investigate how to improve the performance of feature detection using different tools such as Determinants of Covariance Matrices, Gradient Correlation Matrices, Laplace transformation in Gaussian scale space and the evolution difference of planar curves in B-spline scale space. Since corner is one of the important feature information used in these methods, we propose various corner detectors in this paper and evaluate their performance. In this paper, we present our achievements on the following topics:1. Based on the analysis of the eigenvalue and the eigenvector of the covariance matrix, we propose the determinant of the covariance matrices (DCM) algorithm which uses the covariance matrix of the region of support of the planar curve as its response function.2. Besides the DCM algorithm, the characteristics of planar curve gradient relative matrix can also reflect the local features of profiles. Based on the characteristics of planar curve gradient relative matrix, we research the eigenvalue and the eigenvector of the matrix and how do they reflect the local change of the planar curve. Based on the geographical meaning of the eigenvalue and the eigenvector of the gradient correlation matrix, we use it to be the response function of the corner detection algorithm and give the detailed solution. We also prove the feasibility of single cornerΓmodel and END double model theoretically. The experiment results show that our algorithm has good detection ratio, location and anti-interference.3. We define the LoG operator of the planar curve as a vector computation and use it to convolution with the profile coordinate. We observe that the norm of the LoG operator has unique extreme value at the corners of a profile. So we define this norm as the corner response function and propose LoG corner detection algorithm. Experiment results show that this algorithm has good accuracy, robustness, less complexity and easy to implement.4. Followed by the foremensioned detector obtained under the Gaussian scale space, we also research on the characteristics of the evolution of the B-spline scale space under multi-scales. We observe that corner area of the profile evolutes significantly under different scales while the other areas have similar response. So we can use the difference of this evolution to locate corners. We define the corner response function as the DoB norm of the evolution difference of the image profile. Using this way, the characteristics of different scales has been efficiently merged. It not only enhances the response of feature point, but also restrains the impact of noise. DoB algorithm solves the problem of single scale of candidate corners and result in higher detection accuracy. Meanwhile, B-spline helps fast convolution which can speed up the detection. Experiment results show that detection has fast speed, precise location and high accuracy.5. Finally, we evaluate the classical Harris, CSS and Wavelet algorithm with our proposed DCM,GCM,LoG and DoB using the metrics of CCN, ACU and Repeatability. Intensive experiments have been implemented on the test case of image rotation, affine, scale and noise. Results show that our corner detection algorithms have good accuracy which implies its usage in real applications.