International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012 
XXII ISPRS Congress, 25 August - 01 September 2012, Melbourne, Australia 
123 
RESEARCH OF IMAGE MATCHING ALGORITHM 
BASED ON ROTATION VECTOR FIELD 
Duan YanSong 
School of Remote Sensing and Information Engineering, Wuhan University, 129 Luoyu Road, Wuhan 430079, China 
Commission IV, WG IV/3 
KEY WORDS: rotation vector field, rotational invariance, image matching, time efficiency 
ABSTRACT: 
A rotation invariant image matching algorithm is introduced in this paper. The feature descriptor of feature points is calculated in 
polar coordinate system which achieves rotation invariant. The Wallis filter and image edge extraction arc applied to reduce the 
influence of noise and light difference. After constructing the image feature vectors, the potential correspondences are found firstly 
and then the Generalized Hough Transform (GHT) is used to purify the matching result. The experiment results of three data sets 
show that the method is robust to image rotation, time and space efficient and is sufficient to produce matching points which can be 
used as initial values for further accurate image matching. 
1. INTRODUCTION 
The air vortex, side wind, nonuniform wind speed and other 
factors may lead to the consequence of a large rotation angle in 
images of aerial photography and they are inevitable in many 
cases. The subsequent remedy flights caused by these aerial 
photographs that do not meet the specifications and 
requirements would result in time delays and economic losses. 
Thus the solution of automatically large rotation angle aerial 
image matching and related problems has a very important 
significance. Matching methods based on gray correlation in 
conventional image matching of aerial photography have been 
widely used. Pyramid image strategy together with gray 
correlation can meet demands even if the image has small angle 
rotation. However, when the angle of rotation between images 
surpass 15 °, this matching strategy is difficult to get the desired 
results. Considering the large angles of rotation in image 
matching, SIFT (Scale Invariant Feature Transform) operator 
introduced by David G. Low is the most influential algorithm 
currently. Compared with the traditional method based on gray 
scale, SIFT operator has a good rotation invariance and scale 
invariance. However SIFT feature matching is time-consuming 
and the matching accuracy is not that high. The main reason is 
that SIFT feature points have large amount of attributes. The 
time cost would be huge when traversing the characteristics of 
each point. In addition, SIFT operator use the minimum 
Euclidean distance as the similarity measurement, the overall 
rate of correct match is too difficult to improve. Aiming at the 
shortcomings of SIFT algorithm, a variety of methods have 
been proposed to improve it, such as: SURF (Speeded Up 
Robust Features), PCA-SIFT, GLOH (Gradient location- 
orientation histogram). These methods did much improvement 
on the SIFT algorithm, but the time cost and complexity of the 
processing are still very large. 
We re-examine the aerial photography, aviation aircraft has 
greater immunity for rolling and pitching, so the rolling angle 
and pitching angle are generally not significant. The notable 
change exists in the yaw angle. Yaw angle results in the rotation 
of image plane in the image matching. As long as a feature 
descriptor which is not related to rotation (i.e. rotation invariant) 
can be found, the large rotation angle problems in the aerial 
image matching would be solved. 
2. THE MATCHING ALGORITHM BASED ON 
ROTATION VECTOR FIELD 
To establish the rotation invariant features, we create an 
image window which takes the target position as the center of 
the coordinate axis, and then a polar coordinate system is 
established, as showed in Figure 1, we definite the rotation 
vector feature as follows: 
F(r) = f d G(r,0) 
9=0 
Where r is the radius, 0 is the angle of rotation; G is the 
attribute of the inspecting pixels, in this paper it’s the image 
gray value. 
Figure 1 Rotation vector feature 
When the image is rotated, pixels in the radius circle which is 
concentric follow this rotation. F (r) is unchanged. So, F (r) is a 
rotation invariant. 
Based on the inspecting vector F (r), the similarity measure 
function is defined as follows: 
/ = Z|F(r)-A(r)| 
r=0 
Where F (r) is the reference window’s rotation vector, and F 
'(r') is the target window’s rotation vector, when f equals 0, the 
two features correspond.