A Novel Automatic Structural Linear
Feature-based Matching Method Based on New
Concepts of Mathematically-Generated-Points
and Lines
Somayeh Yavari, Mohammad Javad Valadan Zoej, Mahmod Reza Sahebi, and Mehdi Mokhtarzade
Abstract
This paper investigates reliable automatic high resolution im-
age to map matching using a novel structural linear feature-
based matching (
SLIM
) method. The main components used
by this method are the specific patterns as well as the lines
and points generated mathematically. These components are
produced by extension and intersection of extracted line-seg-
ments. Due to the high numbers of extracted line-segments in
both image and object space, the number of possible patterns
is very high. In order to decrease the search space, the inno-
vative
SLIM
method is performed in three main phases. In the
first phase, using a new weighting procedure, only optimum
numbers of high-qualified well-distributed patterns, which are
more likely to have any correspondence in object space, are
selected. In the second phase, the aim is to find a pair with
maximum numbers of conjugate lines. To do so, all the pos-
sible patterns in object space are screened for each selected
image pattern using four predefined geometric criteria. Simul-
taneously, the correspondence of the other crossing lines is
also determined in the same manner. In third phase, the pair
with maximum numbers of matched-lines is selected among
all the results of second phase. Additionally, the final-phase
is done to increase the amount of correctly matched-lines.
The main contribution of this investigation is automatic and
correct matching of linear features with no need to any initial
information. Additionally, the end-points of the correspond-
ing lines are not necessarily conjugate points. The results
show the high potential of the proposed method in terms of
accuracy, reliability, automation, and time reduction even in
images with repetitive patterns or a high numbers of outliers.
Introduction
The image to map/model registration or georeferencing is a
fundamental step in many applications of image processing
and analysis such as mapping, map updating, environmental
monitoring, image fusion, change detection, robotics, medi-
cal image processing, and remote sensing (Goshtasby, 2005;
Zitova and Flusser, 2003).
The registration process is categorized into four steps includ-
ing: feature extraction, feature matching, transformation function
estimation, and resampling (Goshtasby, 2005; Zitova and Flusser,
2003). Feature extraction is a well-studied step, and matching
and transformation function estimation are important steps.
The common matching algorithms are divided to area-
based matching and feature-based matching (Goshtasby, 2005;
Zitova and Flusser, 2003). Area-based matching methods use
intensity values to find the correspondence; therefore, they
are suitable for image to image registration purposes. Feature-
based matching techniques are categorized as: areal/region
feature-based matching (Goncalves, 2011; Goshtasby, 2005;
Jaw and Wu, 2006; Long and Jiao, 2012), linear feature-based
matching (Chen and Shao, 2013; Habib
et al
., 2004; Heuvel,
2003; Karjalainen
et al
., 2006; Ok
et al
., 2012; Tommaselli
and Medeiros, 2010; Wang
et al
., 2008; Wang
et al
., 2009)
and point feature-based matching (Han
et al
., 2012; Hu
et al
.,
2015; Kang
et al
., 2014; Lowe, 2004; Sedaghat
et al
., 2011).
In some cases, the combination of different features are also
used in matching (Long
et al
., 2015; Zhang
et al
., 2011).
Relational/structural matching is an extension of feature-
based methods when the features are studied as primitives and
matching is based on finding the correspondence between struc-
tural descriptors of primitives in two spaces. In these methods,
not only the features, but also their topological and geometrical
relations are used (Vosselman, 1992; Wang, 1998). Regarding that
structural matching methods do not need any initial approxima-
tions or information (Vosselman, 1992; Wang, 1998), they are
suitable when automatic image to map matching is performed.
In traditional photogrammetry, mathematical models are
solved based on point features. Thus, the major part of the
procedure is done manually using an expert operator. Howev-
er, researchers are persuaded to pass from “point photogram-
metry” toward “line photogrammetry” and “generalized point
photogrammetry” (Schenk, 2003; Zhang and Zhang, 2004).
Linear features have unique characterizes such as easier and
more reliable process of extraction, the potential of utiliz-
ing the relations between lines, abundance of linear features
in high resolution images, and permanent existence of these
features such as roads (Habib
et al
., 2004; Schenk, 2004).
Regarding these advantages, linear features are selected in
this paper as the base features for image to map matching.
Selection of an appropriate mathematical model is another
important issue in the matching procedure. In terms of pho-
togrammetry, there are two common mathematical models.
They are parametric models, which reconstruct the image
geometry at the time of imaging (Valadan Zoej and Sadeghian,
2003; Valadan Zoej and Petrie, 1998), and non-parametric
models which are mainly interpolative in nature (Fraser and
Hanley, 2003; Fraser and Yamakawa, 2004; Tao and Hu, 2001;
Teo, 2013). Regarding parametric models, many researchers
Photogrammetry and Remote Sensing Engineering
Department, Geodesy and Geomatics Faculty, K. N. Toosi
University of Technology, 322 West Mirdamad Ave. 19696,
P.O. Box: 15875-4416, Tehran, Iran (
).
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 5, May 2016, pp. 365–376.
0099-1112/16/365–376
© 2016 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.5.365
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2016
365
