An Improved Harris-Affine Invariant Interest Point Detector

Interest point detection is essential process for many computer vision applications, which must provide invariant points to several image variations, such as, rotation, zoom, blur, illumination variation and change of viewpoints. Harris-Affine detector is considered as one of the most effective interest point detectors, although it still presents vulnerability to some image. This paper proposes an improved Harris-affine interest point detector based on two-dimensional weight Atomic functions to improve the repeatability and stability of detected points of Harris-Affine detector, in which the Gaussian kernel is replaced by the ) (x up Atomic Function. Evaluation results show that the new image interest point detector, called Atomic Harris-Affine detector, improves the repeatability in about 40% compared with the conventional Harris-Affine detector, under several conditions such as blurring, illumination changes, change of viewpoint, as well as other geometrical and affine transformations. Key-Words: interest points detector; Atomic function; Harris-Affine detector; repeatability; convergence rate