MEASUREMENT SCIENCE REVIEW, Volume 13, No. 3, 2013
132
CGCI-SIFT: A More Efficient and Compact Representation
of Local Descriptor
Dongliang Su
1
, Jian Wu
1
, Zhiming Cui
1
, Victor S. Sheng
2
, Shengrong Gong
1
1
The Institute of Intelligent Information Processing and Application, Soochow University, Suzhou 215006, China
2
Department of Computer Science, University of Central Arkansas, Conway 72035, USA,
jianwu@suda.edu.cn
This paper proposes a novel invariant local descriptor, a combination of gradient histograms with contrast intensity (CGCI),
for image matching and object recognition. Considering the different contributions of sub-regions inside a local interest region to
an interest point, we divide the local
interest region around the interest point into two main sub-regions: an inner region and a
peripheral region. Then we describe the divided regions with gradient histogram information for the inner region and contrast
intensity information for the peripheral region respectively. The contrast intensity information is defined as intensity difference
between an interest point and other pixels in the local region. Our experimental results demonstrate that the proposed descriptor
performs better than SIFT and its variants PCA-SIFT and SURF with various optical and geometric transformations. It also has
better
matching efficiency than SIFT and its variants PCA-SIFT and SURF, and has the potential to be used in a variety of real-
time applications.
Keywords: Image matching, descriptor, SIFT, CGCI-SIFT, real-time
1.
INTRODUCTION
MAGE MATCHING is a primary technology in computer
vision and image processing. Among the image matching
algorithms, local descriptor algorithms [1] are more stable.
Local descriptors are discriminative, and robust to partial
occlusion. They do not require segmentation preprocessing,
and are invariant under a variety of transformations. All
these properties make local descriptor algorithms be widely
applied in many fields, such as, content-based large-scale
retrieval [2], video analysis, copy detection, object
recognition, photo tourism, and 3D reconstruction [3].
A good local descriptor algorithm should have following
characteristics: no necessity of pre-segmenting images [4],
high repeatability of feature detector, low dimension of
feature descriptor, robustness to partial occlusion, and
invariance against image transformations, such as,
illumination, rotation, scale, blur, and affine.
Local descriptors have received considerable attention in
recent years. Harris proposed the Harris corner detector [5],
based on the eigenvalues of the second-moment matrix, but
it is not scale-invariant. Lowe introduced a scale invariant
feature transformation (SIFT) [1]. It is invariant under a
variety of transformations, such as scale and viewpoint
changes, rotation, and illumination transformations.
Mikolajczyk and Schmid [6] showed that SIFT is one of the
most effective image matching algorithms against viewpoint
and scale transformations. However, the dimensionality of a
SIFT descriptor is high. This results in inefficiency in real-
time applications. In order to improve the matching
accuracy and reduce the matching time, various extensions
of SIFT have been proposed. For example, Ke and
Sukthankar proposed PCA–SIFT [8], which uses image
gradient patch, and applies principal component analysis
(PCA) to replace the smoothed weighted histograms in SIFT
to reduce the size of a descriptor. It performs better on
artificially generated data. E.N. Mortensen proposed GSIFT
[9], which combines SIFT with global texture information.
H. Bay proposed SURF [7], which has similar steps with
SIFT. But SURF adopts a new processing method for each
step. Its computing speed is faster. E. Tola proposed a
descriptor DAISY [10], which computes dense depth and
occlusion maps from wide-baseline image pairs on the basis
of the EM algorithm. It is very efficient for intensive
computing. Yang and Sluzek [11] proposed a low dimension
descriptor combined with shape features and location
information.
Local descriptor algorithms consist of three primary steps.
First, interest points are detected at distinctive locations in
an image, such as corners. Second, the local region of the
interest point is represented by a feature vector. The
descriptor has to be distinctive, robust to noise and detection
errors, and invariant against geometric and photometric
transformations. Finally, vectors of descriptors are matched
between different images. Many extensions of SIFT are
mainly related to the construction of the SIFT descriptor.
The algorithm proposed in this paper is also related to the
improvement of the SIFT descriptor.
In this paper, we propose a novel invariant local descriptor,
a combination of
gradient histograms with contrast intensity
(CGCI), for image matching and object recognition. It
exploits contrast intensity information by evaluating
intensity difference between an interest point and other
pixels in the local region. It is one of the extensions of a
standard descriptor SIFT, called CGCI-SIFT in following
paragraphs. It is more efficient than SIFT and its two
variants (PCA-SIFT and SURF), since it can require less
data to represent a local region. Our experimental results in
Section 4 show that it not only achieves significantly better
performance, but also uses less time in both feature
extraction and image matching, comparing with SIFT and its
two variants.
I
10.2478/ms
-2013-0022
Brought to you by | Soochow University Library
Authenticated
Download Date | 1/12/16 7:09 AM