The following Section, the details of the proposed method are
described. We demonstrate and analyze the experiment results
using diverse datasets from different sensors (
UAV
-borne,
airborne and spaceborne data) under challenging scenarios
such as images of bushes, forests, and crops. The final Section
concludes this paper and provides insight on future work.
Related Work
In general, the existing matching algorithms can be mainly
divided into area-based methods and feature-based methods
(Xiong and Zhang, 2009). In area-based methods, image data
in a rectangular window is used in the form of a matrix of
grey values to describe the centric pixel. Among area-based
methods, cross-correlation is one of the most widely used
similarity measurements. The Normalized Cross-Correlation
(
NCC
) was proposed to increase the robustness of cross-corre-
lation to linear grayscale variation and slight geometric distor-
tion (Helava, 1978; Ackermann, 1984; Lhuillier and Quan,
2002). To improve the matching reliability, some matching
strategies like least squares matching (Gruen, 1985, 2002, and
2005) and relaxation labeling (Wu and Pairman, 1995) were
proposed to constrain the similarity measurement. An advan-
tage of area-based methods is that it can achieve sub-pixel ac-
curacy and even better in some cases but they are sensitive to
image nonlinear intensity change and perspective geometric
deformation (Gruen, 2012).
Feature-based methods can overcome some of the afore-
mentioned problems of area-based methods by exploring
local image information to construct transformation-invariant
descriptors for detected features. Feature-based matching
methods attracted much attention in recent years as the suc-
cess of the well-known Scale Invariant Feature Transform
(
SIFT
) (Lowe, 2004). Without too much derivation, the para-
digm of feature-based methods can be generally divided into
three steps, being feature detection, description and matching.
Existing methods either focus one or two of these three steps,
or address them in a row. The most commonly used detectors
are the Förstner detector (Förstner and Gülch, 1987), Harris-
based detectors (Harris and Stephens, 1988; Mikolajczyk and
Schmid, 2004), Smallest Univalue Segment Assimilating Nu-
cleus (
SUSAN
) detector (Smith and Brady, 1997), Phase Con-
gruency detector (Kovesi, 1999), Difference of Gaussians (
DoG
)
detector (Lowe, 2004) and Maximally Stable Extremal Region
(
MSER
) (Matas
et al
., 2004), as they comparatively show higher
feature repeatability rates in either benchmarks or practical
tests (Mikolajczyk
et al
., 2005; Tuytelaars and Mikolajczyk,
2008). As for descriptors,
SIFT
, Speeded Up Robust Features
(
SURF
) (Bay
et al
., 2008), Histograms of Oriented Gradient
(
HOG
) (Dalal and Triggs, 2005) and
DAISY
(Tola
et al
., 2010)
are among the top due to their robustness and accuracy in
practical scenarios (Mikolajczyk and Schmid, 2005; Gauglitz
et al
., 2011). Different measures like mutual information and
Euclidean distance can be adopted to evaluate feature similar-
ity across different images (Maes
et al
., 1997). The matching
performance can be improved by using some iterative match-
ing strategies (Morel and Yu, 2009; Yu
et al
., 2012; Chen
et
al
., 2013; Chen
et al
., 2017). Feature-based methods in remote
sensing image matching can be specially treated, since the
spatial and geometric information of the often geo-referenced
images can be used to constrain the correspondences search
area. This not only accelerates the matching procedure but
also improves the matching precision (the ratio between the
number of correct matches and the number of total matches).
Despite these well-investigated feature detection and
matching methods generally work in many practical sce-
narios, such seemingly robust methods invariably run into
problems when applying on images with highly repetitive
textures. In order to improve the matching performance in
such scenario, the global context is often explored by com-
puting inter-relationship between correspondences as spatial
constraint (Mortensen
et al
., 2005; Duchenne
et al
., 2011).
Inspired by this idea, methods were proposed to match
repetitive patterns by matching point-pair and then finding
one-to-one correspondences from the matched pairs based
on geometric constraints (Fan
et al
., 2011). Local ambiguities
can be reduced by extending support regions to extract more
information from larger regions. These methods work well
for images with moderate-level of repetitive patterns, while
the performance of matching relies heavily on the initially
matched pairs. If repetitive textures take a large portion of
the image content, the information extracted from extended
regions may not be distinctive enough. Existing methods also
considered scale and orientation constraints provided by
geo-referencing information, e.g., Global Navigation Satellite
System and Inertial Navigation System (
GNSS/INS
) data. These
data can be combined with local features to match images
with large number of similar patterns (Habib
et al
., 2016).
The use of geometric constraints is capable of reducing false
matches to a certain extent, while the feature descriptors may
be still computed from similar texture regions that limit the
matching performance.
To further improve the capability of handling repetitive
patterns, increasing the feature support region size might not
bring as many benefits as the downside of causing higher
chances of geometric deformation (i.e., perspective or affine
distortion) and higher computational complexity. Therefore,
in this paper, we tend to instead explore the texture struc-
tures within a local region by discovering distinctive features
within a texturally repetitive area.
Methods
Our proposed method considers a full pipeline of feature
point detection, description, and matching. Different from
the traditional paradigm which detects features and compute
descriptors in two separate steps, the feature description is
integrated into the feature detection step. As a first step of
the method, a novel
LDF
detector is proposed to measure a
two-level distinctiveness of the pixel and the support region
through a modified response function, and detect
LDF
s with
high matching potential. Seed points are then selected from
the
LDF
s and matched using descriptor similarity through
a bidirectional matching strategy. These seed point match-
ing results are further filtered through a matching reliability
indicator, where only the pair of seed points with the high-
est reliability value in each area is kept. The remaining seed
point matching results are filtered again using the Random
Sample Consensus (
RANSAC
) (Fischler and Bolles, 1981).
Based on these matches, a coarse geometric transformation
(e.g., affine or projective transformation) can be computed
between the source and target image to bound the search area
for more refined matching. Finally, a
FIPS
-based search strat-
egy is proposed. It is combined with descriptor similarity and
geometric constraint to compute correspondence for each
LDF
on the source image. The workflow of the proposed matching
method is summarized in Figure 1.
LDF Detector
Classic corner detectors, for instance Harris detector, use
corner response values computed from covariance matrix to
measure the potential corner pixels as interest points. Since
the covariance matrix is aggregated through a small window,
such corner points can be distinctive in this local window
while being ambiguous if placed under a larger scale (feature
support region to compute descriptor). Such globally non-
distinctive feature points are normally becoming disturbing
514
August 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
