Interest point detection using rank order LoG filter Zhenwei Miao n , Xudong Jiang School of Electrical and Electronic Engineering, Nanyang Technological University, Nanyang Link, Singapore 639798, Singapore article info Article history: Received 2 November 2011 Received in revised form 28 February 2013 Accepted 26 March 2013 Available online 12 April 2013 Keywords: Interest point detection Image matching Weighted rank order filter Repeatability abstract This paper proposes a novel nonlinear filter, named rank order Laplacian of Gaussian (ROLG) filter, based on which a new interest point detector is developed. The ROLG filter is a weighted rank order filter. It is used to detect the image local structures where a significant majority of pixels are brighter or darker than a significant majority of pixels in their corresponding surroundings. Compared to linear filter based detectors, e.g. SIFT detector, the proposed rank order filter based detector is more robust to abrupt variations of images caused by illumination and geometric changes. Experiments on the benchmark databases demonstrate that the proposed ROLG detector achieves superior performance comparing to four state-of-the-art detectors. Evaluation experiments are also conducted on face recognition problems. The results on five face databases further demonstrate that the ROLG detector significantly outperforms the other detectors. & 2013 Elsevier Ltd. All rights reserved. 1. Introduction As a powerful tool for computer vision, interest point detection has drawn great attentions in the last two decades [1–3]. It has been used in a wide range of research, such as panoramic image stitching [4], image retrieval [5], image registration [6], texture classification [7], object categorization [8], object recognition [9,10], 3D object modeling [11], video shot retrieval [12], and face recognition [13]. Many interest point detectors have been pro- posed in the past few years to detect local structures of images [14–26]. They can be roughly classified into three categories: corner-based detectors, blob-based detectors and region detectors. Corners correspond to points in the 2D images with high curvature [3]. Harris corner detector [14] uses the second moment matrix, also called the auto-correlation matrix, to analyze the principal intensity changes in two orthogonal directions in a neighborhood around a point. The Harris measure combines the trace and the determinant of the second moment matrix in a single measure. This measure is used to detect the image local structures where the principal intensity changes in two orthogonal directions are both large. However, this type of structures includes not only corners, but also textured patterns and noise [27]. Harris–Laplace/affine detectors [15] were proposed to be invariant with scale and affine changes. Corners are detected by the Harris corner detector in multi-scales, and their characteristic scales are determined by the Laplacian operator. As the shape of a corner does not match the shape of the Laplacian operator, scale estima- tions for corners are often unstable [28]. SUSAN detector [16] defines a corner as the smallest USAN (univalue segment assim- ilating nucleus) point, which is dissimilar from a majority of pixels within a neighborhood of it. This detector is sensitive to impulse noise and blur, and it fails to deal with scale changes. Blobs refer to bright regions on dark backgrounds or vice versa [29]. Hessian detector [18] employs the Hessian matrix to analyze the second order Taylor expansion of the intensity surface. The Hessian matrix consists of the second order derivatives of image intensity. The trace and the determinant of this matrix are used to detect blobs in a single scale. Hessian–Laplace/affine detectors [15] were developed to detect blobs in multiple scales based on the Hessian detector and the Laplacian operator. These detectors are stable in estimating the characteristic scales of blobs, of which the shapes are similar to that of the Laplacian operator. SIFT detector [19] employs the difference of Gaussian (DoG) filter to approx- imate the normalized Laplacian of Gaussian (LoG) filter. The DoG filter significantly accelerates the computation process. SURF detector [20] employs the box filters and the integral images to further speed up the Hessian–Laplace detector. The box filters are approximations of the second order Gaussian derivative filters. The integral images allow for the fast convolutions of the box filters with the input image. Different approaches, which are not based on the second order derivative of image intensity, were proposed in [21–23] to detect blobs. Salient region detector [21] employs the image local complexity to detect blobs. The characteristic scales are determined by the entropy extrema of the local descriptors. A common computational concept is proposed in [22] to detect different types of local structures. The intensity variance in a local circular region is divided into three components, which are used to Contents lists available at SciVerse ScienceDirect journal homepage: www.elsevier.com/locate/pr Pattern Recognition 0031-3203/$ - see front matter & 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.patcog.2013.03.024 n Corresponding author. Tel.: +65 9220 7858. E-mail addresses: mi0001ei@e.ntu.edu.sg, mi0001ei@ntu.edu.sg (Z. Miao), exdjiang@ntu.edu.sg (X. Jiang). Pattern Recognition 46 (2013) 2890–2901