A Tree-Structured Feature Matching Algorithm

Feature matching is essential in computer vision. In this paper, we propose a robust and reliable image feature matching algorithm. It constructs several matching trees in which nodes correspond to traditional sparsely or densely sampled feature points, a

  • PDF / 2,956,828 Bytes
  • 11 Pages / 439.37 x 666.142 pts Page_size
  • 108 Downloads / 222 Views

DOWNLOAD

REPORT


Hefei Institutes of Physical Science, Chinese Academy of Science, Hefei, China [email protected], [email protected] Department of Computer Engineering, AnHui SanLian University, Hefei, China [email protected] Abstract. Feature matching is essential in computer vision. In this paper, we propose a robust and reliable image feature matching algorithm. It constructs several matching trees in which nodes correspond to traditional sparsely or densely sampled feature points, and feature lines are constructed between the nodes to build a cross-references based on a Difference-of-Gaussians down-sampling pyramid. This can make patchbased descriptors combine efficiently with spatial distributions. By comparing with SIFT, SURF and ORB, our method can get much more correct correspondences on both synthetic and real data under the influence of complex environments or transformations especially in irregular deformation and repeated patterns. Keywords: DoG · Image feature matching · Tree structured matching · Feature line

1

Introduction

Feature correspondence is a fundamental task in many applications of computer vision such as feature tracking[13], image classfication[11], object detection[4], 2D and 3D registration[10,17]. A large number of applications promote various kinds of feature matching algorithms. At the same time, the continuous development of the applications put forward some new and higher requirements such as precision, speed and robust ability. In order to meet the needs of all these practical applications, much attention has been paid to improve the matching performance. A widely used method is computing variety of feature descriptors and select a threshold carefully to filter out large outliers. Therefore, variety of feature descriptors have been proposed, such as SIFT[8], SURF[9], BRISK[15], ORB[14] and LDB[19]. Further more, some people turned to combine more flexible geometric features and spatial characteristics. For instance, Chui et al.[5] introduced a feature based method named TPS-PRM (thin-plate spline-robust point matching) for non-rigid registration. C.Schmid[16] and Y.Zheng[20] use the thought of proximity. They assumed that two adjacent points in the original image should be matched to the couples which are also neighbours in the target image. X.Xu[18] use RANSAC and strong space constraints to obtain relatively stable feature point set first and then use a selection model[10] to decide which transformations are the most appropriate one. Finally, it constructs a global geometric c Springer-Verlag Berlin Heidelberg 2015  H. Zha et al. (Eds.): CCCV 2015, Part II, CCIS 547, pp. 190–200, 2015. DOI: 10.1007/978-3-662-48570-5 19

A Tree-Structured Feature Matching Algorithm

191

transformation model as the matching constraint. O.Duchenne[6] accommodate both (mostly local) geometric invariants and image descriptors and search for correspondences by casting it as a hyper graph matching problem using higher order constraints. Although many existing algorithms are general and could cover both rigid and non-ri