A Scale-Invariant Keypoint Detector in Log-Polar Space
Tao Tao, Yun Zhang
College of Information Engineering and Automation, Kunming University of Science and
Technology, Kunming, Yunnan, P. R. China
ABSTRACT
The scale-invariant feature transform (SIFT) algorithm is devised to detect keypoints via the difference of Gaussian
(DoG) images. However, the DoG data lacks the high-frequency information, which can lead to a performance drop of
the algorithm. To address this issue, this paper proposes a novel log-polar feature detector (LPFD) to detect scale-
invariant blubs (keypoints) in log-polar space, which, in contrast, can retain all the image information. The algorithm
consists of three components, viz. keypoint detection, descriptor extraction and descriptor matching. Besides, the
algorithm is evaluated in detecting keypoints from the INRIA dataset by comparing with the SIFT algorithm and one of
its fast versions, the speed up robust features (SURF) algorithm in terms of three performance measures, viz.
correspondences, repeatability, correct matches and matching score.
Keywords: Image matching, log-polar space, scale invariance, keypoint.
1. INTRODUCTION
Image matching is an essential task in image processing. It has also found a wide range of applications such as object
recognition, medical diagnosis and weather forecasting [1]. Image matching, in general, consists of four components:
feature space, similarity measure, search space and search strategy [2]. There are two general approaches to image
matching. One is intensity-based algorithm where an image patch is described by its intensity content. The other is
feature-based algorithm where a patch is described by its photometric/geometric feature points and contours.
There are three popular types of similarity measures for intensity-based algorithms, i.e., the mutual information (MI)
[3], the sum of squared differences (SSD) [4] and the normalized cross-correlation (NCC) [5]. The MI algorithm is
designed for alignment of multimodality medical images, where the misalignment is small. The SSD algorithm is often
applied to images without apparent intensity variations, while the NCC algorithm is invariant to affine lighting changes.
On the other hand, the feature-based algorithm has recently attracted more attention for its strong competiveness in
handling images with large appearance changes. In this respect, many methods have been proposed to extract different
types of features from images. For example, the Harris corner detector [6] is devised to detect rotation-invariant corners
with comparable gradient responses in two independent directions. To distinguish a corner from an edge, it also provides
a cornerness measure for a feature. However, the detection is sensitive to the scale changes of image. Later, the scale-
invariant feature transform (SIFT) method [7] is used to detect scale-invariant blubs (or keypoints) with a clear radial
gradient response. It also provides a distinctive descriptor for each feature. Moreover, the speed up robust features
(SURF) method [8] is put forward to reduce the time overhead for detection using integral images. To reach a
comparable performance, the SURF method is about three times faster than the SIFT method.
The SIFT method extracts keypoints from images in scale space, which causes the loss of high-frequency image
information and, thus, leading to a performance drop of the method. To address this issue, this paper proposes a novel the
log-polar feature detector (LPFD) to detect scale-invariant blubs (keypoints) from images in log-polar space, which, in
contrast, can retain all the information the images contain. That method consists of three components, viz. keypoint
detection, descriptor extraction and descriptor matching. Finally, the method is evaluated in extracting keypoints from
INRIA dataset [9] by comparing with the SIFT method as well as its fast version, the SURF method in terms of three
performance measures, viz. correspondences, repeatability, correct matches and matching score.
The rest of this paper is organized as follows. First, three components of the LPFD method, keypoint detection,
descriptor extraction and descriptor matching, are presented in Section 2, 3, and 4 respectively. Then, the LPFD method
is experimentally evaluated in Section 5. Finally, our conclusion is given in Section 6.
Eighth International Conference on Graphic and Image Processing (ICGIP 2016), edited by Tuan D. Pham, Vit Vozenilek,
Zhu Zeng, Proc. of SPIE Vol. 10225, 102250P · © 2017 SPIE · CCC code: 0277-786X/17/$18 · doi: 10.1117/12.2267122
Proc. of SPIE Vol. 10225 102250P-1
Downloaded From: http://proceedings.spiedigitallibrary.org/ on 02/23/2017 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx