CS-FREAK: An Improved Binary Descriptor - Advances in Image and Graphics Technologies - page 129

CS-FREAK: An Improved Binary Descriptor

Jianyong Wang, Xuemei Wang, Xiaogang Yang, and Aigang Zhao

Department of Automation, Xi'an Institution of High-Tech
Xi'an, China
[e-mail]

Abstract. A large number of vision applications rely on matching key points
across images, its main problem is to find a fast and robust key point descriptor
and a matching strategy. This paper presents a two-step matching strategy based
on voting and an improved binary descriptor CS-FREAK by adding the neighbor-
hood intensity information of the sampling points to the FREAK descriptor. This
method divides the matching task into two steps, firstly simplify the FREAK[1]
8-layer retina model to a 5-layer one and construct a binary descriptor, secondly
encode the neighborhood intensity information of the center symmetry sampling
points, and then create a 16-dimentional histogram according to a pre-constructed
index table, which is the basis for voting strategy. This two-step matching strategy
can improve learning efficiency meanwhile enhance the descriptor identification
ability, and improve the matching accuracy. Experimental results show that the
accuracy of the matching method is superior to SIFT and FREAK.

Keywords: Point matching, binary descriptor, two-step matching strategy,
FREAK.

1 Introduction

Image matching plays a key role in computer vision, image stitching, target recogni-
tion and other fields [3][4]. Image matching consists of three steps: feature points
detect, feature points description and matching.
After obtaining the feature points, adding an appropriate description is a critical
work, which greatly affects the subsequent efficiency of image matching. Lowe DG
[2] et al. proved that SIFT descriptor is more stable through conducting a performance
evaluation experiments, in which the images were processed by changing blur, light,
scale and a certain perspective transformation. The experiment results show that SIFT
descriptor can get better matching results compared with other descriptors including
shape context information, complex filtering, invariant moments [5], etc. But there
exist shortcomings: the descriptors are made of 128-dimensional vector which is too
high, a large number of feature points got involved in matching, and the search mea-
surements are relatively time-consuming. Therefore, the subsequent emergence of a
variety of descriptors are improvements of SIFT descriptors, such as the PCA-SIFT
using principal component analysis to reduce the dimensionality [6], GLOH descrip-
tors using the log-polar grid interval instead of grid interval [7], SURF descriptor