Accurate Affine Invariant Image Matching
Using Oriented Least Square
Amin Sedaghat and Hamid Ebadi
Abstract
Image matching is a vital process for many photogrammetric
and remote sensing applications such as image registra-
tion and aerial triangulation. In this paper, an accurate
affine invariant image matching approach is presented. The
proposed approach consists of three main steps. In the first
step, two affine invariant feature detectors, including
MSER
and Harris-Affine features are applied for feature extraction.
In the second step, initial corresponding features are selected
using Euclidean distance between feature descriptors, fol-
lowed by a consistency check process. Finally to overcome
low positional accuracy of the local affine feature, an ad-
vanced version of the least square matching (
LSM
) namely,
Oriented Least Square Matching (
OLSM
) is developed. Well-
known
LSM
method has been widely accepted as one of the
most accurate methods to obtain high reliable corresponding
points from a stereo image pair. However, it is sensitive to
significant geometric distortion and requires very good initial
approximation. In the proposed
OLSM
method, shape and size
of the matching window are appropriately approximated us-
ing obtained affine shape information of the initial elliptical
feature pairs. The proposed method was successfully applied
for matching various synthetic and real close range and satel-
lite images. Results demonstrate its accuracy and capability
compared to standard
LSM
method.
Introduction
Image matching is generally defined as the establishment of
correspondence between two or more images of the same
scene and is probably the most important process in digital
photogrammetry (Gruen, 2012). This process is a vital step
in various photogrammetry and remote sensing applications,
such as image registration (Eikvil
et al.
, 2009; Song
et al.
,
2014; Wu
et al.
, 2012), change detection (Duro
et al.
, 2013; Li
et al.
, 2009; Qin and Gruen, 2014; Tian
et al.
, 2014), image fu-
sion (Li
et al.
, 2012; Yang
et al.
, 2011), 3D reconstruction and
mapping sciences (Cheng
et al.
, 2011; Fraser, 2013; Höhle,
2009; Remondino
et al.
, 2014; Tack
et al.
, 2012).
The existing image matching methods can be classified as
area-based matching (
ABM
) and feature-based matching (
FBM
)
approaches.
ABM
techniques use various similarity measures
between the pixel intensities of the identical image patches
to determine image correspondence. The common similarity
measures are cross correlation, mutual information (Gong
et
al.
, 2013), and least squares matching (
LSM
) (Gruen, 1985).
On the other hand,
FBM
methods first extract the salient
features (e.g., points, lines, regions) from the images and then
establish the correspondence between the extracted features.
Harris (Harris and Stephens, 1988), scale-invariant feature
transform (
SIFT
) (Lowe, 2004), speeded-up robust features
(
SURF
) (Bay
et al.
, 2008), and maximally stable extremal re-
gions (
MSER
) (Matas
et al.
, 2004) are some well-known feature
detectors. Although most
ABM
methods, particularly
LSM
,
have exceptionally high positional accuracies, they still re-
quire good initial approximations to assure convergence, and
are sensitive to both significant geometric and illumination
changes. Compared with
ABM
,
FBM
methods are more robust
and deal better with complex geometrical transformations
(e.g., large scale, rotation, and viewpoint differences; but their
matching precisions are not as high as that of
ABM
.
Local invariant features, especially affine invariant fea-
tures, have been shown to be well suited for image matching
due to the fact that they are robust to geometric and illumi-
nation differences. The most popular local invariant feature
algorithm is
SIFT
(Lowe, 2004).
SIFT
algorithm extracts blob
like circular features, so it is not an affine invariant feature
detector. The repeatability of the extracted
SIFT
features
significantly decreases when the viewpoint change between
two images is greater than about 40 degrees. Various affine in-
variant feature methods have been proposed in the literature
(Aanæs
et al.
, 2011; Barandiaran
et al.
, 2013; Gauglitz
et al.
,
2011; Mikolajczyk
et al
., 2005, Tuytelaars and Mikolajczyk,
2008). For most of the local affine invariant feature detectors,
the output shape is an ellipse (Mikolajczyk
et al.
, 2005). A
comparative study of several affine invariant feature detec-
tors, including Harris-Affine, Hessian-Affine (Mikolajczyk and
Schmid, 2004),
MSER
(Matas
et al.
, 2004),
IBR
,
EBR
(Tuytelaars
and Van Gool 2004), and Salient Regions (Kadir
et al.
, 2004),
was presented by Mikolajczyk
et al
., (2005). Based on their
results the highest score was obtained by the
MSER
detector,
followed by Hessian and Harris-Affine detector.
Recently, local invariant feature detectors and descriptors
as an advanced type of
FBM
methods in computer vision have
been widely applied in photogrammetry and remote sens-
ing image matching and registration. For example, Cheng
et
al
., (2014) proposed an automatic registration approach for
remotely sensed images acquired in coastal areas based on
MSER
(Matas
et al.
, 2004) algorithm and
RANSAC
(Fischler and
Bolles, 1981). Other approaches based on local features were
presented in (Han
et al.
, 2012; Joglekar
et al
., 2014; Kang
et
al.
, 2014; Sedaghat
et al.
, 2012; Sedaghat and Ebadi 2015a;
Sedaghat and Ebadi 2015b; Wu
et al.
, 2011).
Despite the outstanding properties and the considerable uti-
lization of the local invariant feature matching algorithms, they
have a serious problem. The main drawback of the existing
affine invariant feature matching methods, especially in wide-
baseline images, is their limitation in establishment of high
accurate feature correspondence (Mikolajczyk
et al.
, 2005).
Geometric shapes of the affine invariant features are ellipses or
irregular regions and the localization accuracy of the extracted
features is not very high. For this reason, the accurate image
matching using affine invariant features is a difficult task.
Faculty of Geodesy and Geomatics Engineering, K.N.Toosi
University of Technology, No. 1346, Vali-Asr Street, Mirdamad
Cross, Tehran, Iran, 1996715433 (
).
Photogrammetric Engineering & Remote Sensing
Vol. 81, No. 9, September 2015, pp. 733–743.
0099-1112/15/733–743
© 2015 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.81.9.733
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2015
733
