| Corners are local features of images which are the least vulnerable to noise and light change ,geometric transformation, scale transformation and so on, so corner detection has become the key foundation work in computer vision and image processing, also has been widely used in object tracking, scene analysis, image registration and matching, robot navigation,shape representation etc. How to introduce new theories and ideas to improve corner detectors,and how to design new corner detectors which are robust to noises are now being hot problems in image processing.Based on researching on varies classical corner detecting algorithms, this paper is made of several matters following:1)In this paper Gaussian scale space in the traditional linear structure tensors is taken place of B-spline multi-scale space, because the computational complexity of B-spline only depends on the signal or image data itself, but is independent with the scale of B-spline function. Otherwise, B-spline function inherits most of the nice properties of Gaussian kernel, so B-splines are good approximations to the Gaussian kernel.2) In this paper the multi-scale representation of linear structure tensor of image contours is defined in the framework of B-spline scale space. The determinant of the tensor matrix can reflect corners of images as well as possible; then we define the multiplication of determinants of the tensor matrixes as corners'response. True corners are the points whose responses are local extremes of the product when the values of the product exceed a pre-threshold. Finally, the experimental results illustrate our proposed algorithm could both obtain all corners of mostly simple images and be rotation invariant, scale invariant and robustness against noise for real images.The multi-scale product can make the response of corners at different scales preserved, and can effectively suppress noise. Finally experimental results show that the algorithm has good advantages. |
PDF Full Text Request