An Improved SIFT Descriptor Based on In-Out Region Division
Chunlong Xia
Institute of Artificial Intelligence and Robotics
Xi’an Jiaotong University
Xi’an, China
State Key Laboratory of Mathematical Engineering and
Advanced Computing
Wuxi, China
e-mail: swtjuclxia@163.com
Ping Wei
Institute of Artificial Intelligence and Robotics
Xi’an Jiaotong University
Xi’an, China
e-mail: pingwei@xjtu.edu.cn
Abstract—SIFT descriptor is one of the most widely-used local
features for image matching. This paper presents an in-out
region division method to improve the matching rate and the
computational efficiency of SIFT descriptor. The descriptor
region composed of pixel blocks around each keypoint is
divided into the inner and the outer regions, in which different
weights are assigned to the feature vectors. In each pixel block,
a low dimensional feature vector is extracted. A bidirectional
verification strategy is adopted to determine the keypoint
matching. The experimental results show that the proposed
method greatly improves the matching rate and the
computational efficiency.
Keywords-SIFT; region division; bidrectional verification
I. INTRODUCTION
SIFT descriptor [1] is one of the most widely-used local
image features in image processing, computer vision, and
multimedia. For the invariance to image rotation, scaling,
and translation, it plays a crucial role in many applications,
such as image matching, object recognition, and scene
reconstruction.
The SIFT method [1] defines a region of pixel blocks
around the keypoint to compute the feature descriptor, and in
each block, an orientation vector weighted by a Gaussian
parameter is extracted. An image often has plenty of SIFT
keypoints and each keypoint has many descriptor blocks.
The high-dimensional block feature vector leads to the high
dimension of the SIFT descriptor, and thus influences the
computational efficiency. Moreover, the Gaussian weights
applied to the block vector may blur some important image
information, and reduce the discrimination ability of the
descriptor. Thus, reducing the dimension of the block feature
vector and increasing its discrimination ability would be
effective strategies to improve the performance of the SIFT
descriptor.
In this paper, we propose a new method to compute the
feature descriptor of image keypoints. In an image, we detect
multi-scale keypoints, determine their orientations, and
define the pixel blocks around each keypoint with the same
strategy proposed in SIFT [1]. Different from SIFT, we
divide the pixel block region around each keypoint into inner
and outer regions, in which different weights are applied to
the block feature vectors. In each block, a low dimensional
description vector is computed. These strategies reduce the
descriptor dimension and enhance the discrimination ability
of the descriptor. For descriptor matching, we adopt a bi-
directional verification method [2] to compute the matching
keypoint pairs.
We test our method on images with illumination changes,
scale changes, and rotations. The experimental results
demonstrate that our method improves the matching rate and
the computational efficiency.
II. R
ELATED WORK
Local image features have been intensively studied for
the crucial roles in many applications [3]-[5], such as Harris
[6], SIFT [1], SURF [7], ORB [8], and GLOH [9]. Among
them, SIFT descriptor [1] has been widely used for its
invariance to image rotations, scale changes and translations.
The SIFT method [1] includes four general procedures:
detecting keypoints in scale space, determining the
keypoints’ orientations, computing the keypoint descriptors,
and matching the keypoint descriptors.
SIFT descriptors characterize abundant local image
information and thus achieve impressive performance in
many applications. However, the high dimension of the
feature descriptor results in expensive computational cost
and low real-time performance, especially in tasks with
large-scale images or videos. Thus, many studies have
proposed new strategies to improve the original SIFT. PCA-
SIFT [10] reduces the descriptor dimension with PCA
analysis. Some studies compute the descriptor in a circular
region instead of the original rectangular region [9], [11],
[12]. Reference [11] proposes a simplified SIFT descriptor. It
computes the descriptor in a circular region with four
concentric circles and each sub-region defines ten directions.
This method improves the computational efficiency of image
matching. Reference [12] computes the descriptor in a
circular region with eight concentric circles. It defines the
descriptor with sub-region’s gray values and the gray
differences between sub-regions. This method simplifies the
descriptor generation process and improves the
computational efficiency.
2017 IEEE 2nd International Conference on Signal and Image Processing
978-1-5386-0969-9/17/$31.00 ©2017 IEEE