Detection,Description And Matching Of Image Feature

Image feature detection and description are fundamental tasks of many image procession and computer vision. The performance of feature detectors and descriptors directly determines the efficiency and precision of image procession. It is important that image local feature should be as distinctive as possible while also be robust to occlusion, background clutter and noise, invariant to various image transformations due to translation,rotation,scale,affine deformation, difference in illumination, object movement, and change in viewpoint. To these difficulties, this paper is made several matters following:Traditional Gaussian-derived scale space was substituted by B-spline derived scale-space. Convolution algorithm of B-spline function has an efficiency solution because of the specialty of the way B-spline convolved and the specialty of discrete B-spline of order zero. Computation complexity of the convolution algorithm has nothing to do with the scale of B-spline function, but only depends on the signal or image data itself. B-splines are good approximations of the Gaussian kernel, so B-spline derived scale-space inherits most of the nice properties of the Gaussian-derived scale-space. The multi-scale representation of covariance matrix for planar curve over its region of support was defined in the framework of B-spline scale space. Covariance matrix was expanded to scale space, which reduces the influence of computation of covariance matrix caused by noise. The eigenvector corresponding to the largest eigenvalue of covariance matrix indicates contour tangent orientation. Multi-scale space product was introduced into corner detection. Multi-scale space response function of cornerness was defined. The multiplication of change rate of contour tangent orientation at different scales was defined as multi-scale product. As scale becoming larger, the response of corner of different scales was remained. A local extreme of the product was reported as corner when the value of the product exceeded a threshold. Systematic evaluation of the proposed corner detection algorithm was made in this paper. Experiments demonstrated that our algorithm is rotation-invariant and insensitive to slight scale transform. Moreover, this algorithm was compared with other classic detectors, and the experimental results also showed that the new algorithm has good detection, localization performance and good efficiency.Corners in an image were detected by multi-scale Harris operator, and they were taken as initial interest points. Since adaptive non-maximal suppression eliminated lots of potential matching points, condition theory was applied to control the number of initial interest points. The bad conditioned points were eliminated, so the computational complexity of the following process was decreased and the efficiency of the algorithm was improved. At the mean time, matching points were mostly kept. Experimental results showed that the efficiency is improved significantly, the proposed algorithm is good at feature matching, and it has good robustness to geometrical transformations, image noise and illumination change.