This paper proposes a new interest point detection method by using contour shape and the direction derivative information entropy of surrounding pixels. The proposed method can improve localization accuracy and noise robustness. First

the multi-direction imaginary parts of Gabor filters are used to extract the gray variation information of input images to acquire the second small direction derivative. Scale multiplications are used as the measure to detect interest points. The proposed method is based on contour-based methods and is improved in two ways: the information of edge shape and the gray variation are combined to detect the interest points; the sensitivity of scale multiplication to local variation and noise on the edge contour is reduced. Second

the Canny edge detector is used to extract edge maps and fill gaps to obtain the edge contour. Third

the second small direction derivative information entropy of the edge contour pixel and its surrounding pixels is computed. The normalized entropy is used as new corner measure by comparing with the threshold. Finally

non-maximum suppression is applied to the result obtained in the third step. A local window (9×9) is used to slide through all the pixels of the edge contour. If the center pixel of the window is the local maxima within the window

the central pixel scale product is retained; otherwise

the center pixel will be set to zero. Compared with the gray method

which analyzes the contour shape

or the curvature-based method

the proposed method combines the idea of two algorithms that use the gradient direction entropy of the contour pixel and its neighbor pixels as the corner measure. Furthermore

unlike the gradient direction variation in homogeneous regions and edge lines

the information variation on the interest point presents an anisotropic feature. The information entropy according to the second small direction derivative (the first small direction derivative may be zero) is used as a new corner measure to improve localization accuracy. The proposed method has an average score of 1.625 in the performance index

which is significantly higher than the method of Harris (3.25) and He and Yung (2.625) and the CSS (2.5) interest point detection operator. Compared with the three state-of-art algorithms

our approach is competitive with respect to the repetitive rate of interest point

detection accuracy

and noise robustness.