Automatic Registration Method fo
r
Leather Section Image using
SIFT and Wavelet Transform
Xin Zhang, Jinyong Cheng and Huayong Zhang
School of Information, Qilu University of Technology, Jinan, 250353, China
Keywords: Image registration; Key point; Wavelet transform.
Abstract: Image registration is one of the hot topics of image processing, which has been widely concerned by
researchers. However, most existing image registration methods are inefficient and inaccuracy. Thus, this
paper proposes an efficient method that uses multi-resolution wavelet transform to process the original
images. This method utilizes decomposition and reconstruction of two-layered wavelet to eliminate the error
matching points effectively. Experimental results show that this method has a better robustness, which can
not only get more key points but also improve the correct matching rate greatly.
1 INTRODUCTION
Leather is regarded as high-grade consumer goods,
because of its excellent material and attractive
appearance, but its three-dimensional structure of the
fiber, has not been fully discovered. As Nature
starting to crack the mystery of toughness of spider
silk from spider silk structure in literature (Glareh
Askarieh, My Hedhammar and Kerstin Nordling,
2010), it provides guiding significance for the
manufacture of man-made fibers manufacture.
With the development of image processing
technology, much attention has been paid to high
resolution, large image processing technology.
However, due to the order of magnitude of a piece of
leather fiber is very small, So it need to enlarge the
larger multiples. If you want to get a complete high-
definition images, image stitching (Koiolovoci,
Vagvolgyib, 2000) in which the technology is
particularly important where the image registration
is the key part of it. Image registration is found the
relative position of point to point and mapping
relation between different images of the same scene
shot, or establish connection to some feature points.
The process usually includes: feature detection,
feature matching, transformation model parameter
estimation, image resampling and transformation. In
recent years, more widely used based on feature
points matching (Thornburg Jonathan, Rodes
Charles and Lamvik Michael, 2006) are SIFT
feature matching algorithm is put forward by Lowe
DG in 2004 (HeMY, DaiYC and ZhangJ, 2008) and
SURF feature matching (Bay H, Tuytelaars T and
VanGool L, 2006) by Bay. SIFT algorithm under the
condition of the scale and rotation changes better
than the SURF (Huang L, Chen C, Shen H, 2015), in
the field of image registration and stitching been
more widely used (Jiachang Gong, Jichang Guo,
2016 and Chen Y, Shang L, 2016).
2 IMAGE ACQUISITION
2.1 Leather interface of image retrieval
The texture of leather fibre is very soft. They usually
occurred deformation easily when we cut the leather
into slices according to the spacing of a few
microns. And it also will cause some changes in
different degrees with the position and shape for
every fibre in the image.
But the leather is porous material, epoxy resin is
infiltrated into leather fiber. The resin will be
infiltration of leather fiber as soon as the leather
touched the resin which is low viscosity and strong
permeability. The leather fibre is fixed with the resin
fixative. Then we can get leather - resin composite
material with a certainty hardness. And the material
can also be mosaic, grinding, polishing and soon
technology.
53
Zhang X., Zhang H. and Cheng J.
Automatic Registration Method for Leather Section Image using SIFT and Wavelet Transform.
DOI: 10.5220/0006443600530057
In ISME 2016 - Information Science and Management Engineering IV (ISME 2016), pages 53-57
ISBN: 978-989-758-208-0
Copyright
c
2016 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
53
2.2 Maintaining the Integrity of the
Specifications
After we get the picture of metallographic sample
preparation, we can confirm the interlamllar spacing
of sequence diagram by measuring the thickness of
the substrate. The leather sample would be put into
the temperature of 40 for one week. After that,
polish the metallographic sample preparation which
is cylindrical to be smooth. Choose four points
which is near the leather sample, then use micron
micrometer to measure the thickness of the four
points and record the dates. We can get the leather
cross section images which are the first of the
sample sequence images via the microscope
observation.
Cut out the leather sample piece to meet the
requirements of the thickness which complete the
operation of take a photo, then take it for polishing,
respectively measuring the thickness of the
cylindrical sample piece of four points use micron
micrometer, and obtain the second image on
microscope, repeat the steps to get images in turn, so
that we can get sequence images. Through
calculation the 4 points thickness before and after in
twice, take the average as the distance between two
image, so we can obtain the layer spacing of
sequence image.
3 SIFT ALGORITHM
3.1 SIFT algorithm description
SIFT algorithm is an operator which is based on
scale space, image scaling, rotation invariance, or
even affine transformation, the noise also maintained
a certain degree of stability of local image features
described. It can detect all of SIFT key points and
128-dimensional feature vectors from each section
of the leather image. By the determination of the
position, size and orientation, it establishment a
descriptor for each key point, described through a set
of vectors to described. In order to achieve image
registration. It can be divided intoeveral steps:
1) Construction scale space, detect extreme
points, to obtain scale invariance.
The problem of SIFT algorithm can be classified as
find the key point in different scale space (Tony
Lindeberg, 1994), and the Gauss kernel function is
the only one that can generate the multi-scale space.
Convolution of Two-dimensional gaussian
function
),,(
σ
yxG
, can achieve the purpose of
blurred image Template Rogos dimensions m*n
(generally speaking, the size is
16*16
+
σ
σ
,
then the image position (x, y) on the template of the
corresponding gauss formula is
2
22
2
)2/()2/(
2
2
1
),(G
σ
πσ
nymx
eyx
=
(1)
The scale space of two-dimensional image
),,(
σ
yxL
can be defined as the convolution
operation of a change in scale two-dimensional
Gauss's function
),,(
σ
yxG
with the original image
),( yxI
:
),(*),,(),,( yxIyxGyxL
σ
σ
(2)
Where * denotes the convolution operation,
σ
is the
scale space factor, the smaller value the less
smoothed, the corresponding scale smaller.
Scale space is realized by Gauss Pyramid, each
pixel in the dog scale space need to compare with
the pixel point of neighboring points around the 8
adjacent points and the upper and lower scales
corresponding to the location of the surrounding
neighborhood 9*2 points a total of 26 pixels, so that
can detect the local minima both in the scale space
and two-dimensional image space.
2) Filter and precise positioning feature points.
DoG operator can generate strong edge response,
and it is needed to eliminate the unstable edge
response points. A poor DoG function at the point of
the peak points have larger principal curvature in the
direction of across the edge, but in the direction of
the vertical edge of is smaller. The relationship of
main curvature and characteristic value is
proportional, so it can be used to determine a
threshold value to eliminate the edge response
points, which can enhance the stability of the
matching and improve the ability of anti-noise.
3) Distribution direction value for the feature
points.
We can use the distribution features of the gradient
direction. We need to specify the direction of
parameter for every key point, then descriptor will
have rotation invariance. For the key points detected
in the DoG pyramid, gather the gradient and
directional distribution feature for pixels which are
in the Gaussian pyramid image neighboring window.
ISME 2016 - Information Science and Management Engineering IV
54
ISME 2016 - International Conference on Information System and Management Engineering
54
Below is the module value of gradient and
directional for the pixels
22
))1,()1,(()),1(),1((),( +++= yxLyxLyxLyxLyxm
(3)
))
)
,1(),1(/))1,()1,(((tan,
1
yxLyxLyxLyxLyx ++=
θ
(4)
where L represents value of a key point in spatial
scale.
Direction of sampling points is carried out by the
adjacent area window with the key points as the
center. It shown through 36 columns (Divides the 0-
360 to 10 degrees for a column).The direction of the
key points is the peak of the histogram.
4) Generate feature descriptor
After the calculated of characteristics of the key
points through the above steps, The value of the key
points feature can be determined, we can describe
the key points through a set of vectors, so that it is
not affected by changes in perspective, the impact of
changes in perspective.
First, determine the calculation the image area of
descriptor. SIFT descriptor is a representation of
statistics results about the key point adjacent area of
Gaussian image gradient, In (David G.Lowe, 2004)
Lowe papers suggest that descriptor use eight
directions information calculated by 4*4 window in
the key points scale space, in total4*4*8=128
dimensional vector. Descriptor gradient direction
histogram are computed by the blurred image of the
scale where the key point exist, the radius of image
region is expressed as:
2
1)1(*2*3 ++
=
d
radius
oct
σ
(5)
where
oct
σ
indicates the scale of group containing
key point.
Then rotate the coordinate axis as the key point
of the main direction, so the new coordinates after
the rotation is:
=
y
x
y
x
θθ
θθ
cossin
sincos
'
'
]),[,( radiusradiusyx
(6)
The last, Calculate gradient value of each pixel
which in the area of image radius and distribute it to
eight directions, calculate the weights. Setting
threshold so that we can truncation larger gradient
value, then normalized processing again, remove
some direction of gradient value which is the value
of gradient too large cause by the illumination
change to improve the identification of features. If
ratio of the nearest two key point distance to the
nearer two key point distance is less than the
proportion of threshold then accept the matching
points.
3.2 Principle of Wavelet Transform
Wavelet transform is a major breakthrough of
Fourier transform and short time Fourier transform,
and it is successfully applied to image denoising,
edge detection and so on. Compared with the Fourier
transform, The signal can be local subdivided by
wavelet transform in the two dimensions of time and
frequency.
When we use wavelet analysis the image, we
should choose the appropriate basic wavelet or
wavelet, wavelet function is formed by a series of
basic wavelet translation, flexible operation, and
finally the signal will be projected onto the signal
space composed of translation and stretch wavelet
for analysis.
1) Wavelet definition
Set
)()(
2
RLx
ϕ
, and meet the conditions
00 =
ϕ
,
then we called
)(x
ϕ
as basic wavelet or wavelet, the
wavelet function is obtained by scaling and
translation:
)(
1
)(
,
a
bx
a
x
ba
=
ϕϕ

0,, aRba
(7)
Where a is stretch factor, b is the translation factor.
The continuous wavelet transform of any
function
)()(
2
RLxf
is defined as:
dx
a
bx
xf
a
fbafW
Rba
)()(
1
,),)((
,
>==<
ϕϕ
ϕ
(8)
The local change of the continuous wavelet
transform, the high frequency resolution, low
resolution in low frequency, which is more
conducive to the extraction of features in the image.
2) Image registration based on wavelet
transform
Wavelet analysis has multi-scale characteristics, it
provides a more flexible processing method for
image processing. In this paper, we use the discrete
Automatic Registration Method for Leather Section Image using SIFT and Wavelet Transform
55
Automatic Registration Method for Leather Section Image using SIFT and Wavelet Transform
55
wavelet transform, the image is decomposed into 2
layers, and the low frequency components of the 2
layer are used to match the image. Experiments
show that the low frequency component contains
most of the information of the image, through the
removal of high frequency components of noise,
improve the robustness of the matching.
DbN wavelet is the discrete orthogonal wavelet
designed by Daubechies, generally referred to as
dbN, where N is the order of the wavelet, the size of
the N reflects the smoothness and concentration of
the wavelet. Only N = 1 (Haar wavelet) has an
analytic expression. Haar wavelet as the wavelet
analysis of the most typical wavelet, the analytical
method of the expression as follows:

other
x
x
x 11/2
1/20
0
1
1
<
<
=
ϕ
(9)
The corresponding scaling function is:

other
10
0
1 <
=
x
x
ϕ
(10)
DbN wavelet function and the effective support area
of the scale function is 2N-1, and the vanishing
moment of wavelet function is N, so its scalability is
better, which can effectively solve the boundary
problem.
The image processing algorithm based on
wavelet design as follows:
Firstly, use multi-scale wavelet analyze and
deal with the image which have much noise,
we can get low frequency images by the 1
layer wavelet decomposition;
Using the obtained image to carry out wavelet
threshold denoising operation, after that,carry
out the operation of 2 layer decomposition,
then we can obtain the two stage low
frequency image;
Reconstruction of two stage low frequency
images and use it for registration;
4 EXPERIMENTAL RESULTS
The experiment done in matlab programming under
Windows 7 in this article, the experimental data is
image slices of the pigskin leather, on the condition
of leather image translation, rotation, and the change
of illumination to complete the operation of image
registration, the result as follows.
Table 1: Comparison of image registration.
Image
key point
count (1
st
image)
key point
count (2
nd
image)
Matching
point
Correct
rate
Original
image
1278 1663 5 80%
Wavele
tImage
1404 1763 7 100%
Image registration is affected by illumination,
position, etc. In this experiment, we use different
layers of images, there are differences in the position
and brightness, and the edge of the feature area is
not obvious. The matching results of the original
image are shown in Figure 1, Figure 2.
Figure 1: Original image a,b
Figure 2: The results of original image registration
The image matching results of the after 2 layers
of wavelet transform are shown in Figure 3, Figure 4.
ISME 2016 - Information Science and Management Engineering IV
56
ISME 2016 - International Conference on Information System and Management Engineering
56
Figure 3: The image after wavelet transform a, b
Figure 4: The results of the image after wavelet transform
Experimental results show that the original
image matching have error point, but using wavelet
to low-frequency image processing, it not only
makes the image edge response is more sensitive,
but also through the noise processing, reduce the
matching error, correct matching rate increased
significantly. Compared to the original image, the
image after wavelet transform can get more
matching point, use the method, iamge registration
can not only improve efficiency, and has better
robustness and higher accuracy.
5 CONCLUSION
This experiment use leather image in the article,
Experiment results show that the image after wavelet
image processing, the feature points matching rate of
SIFT algorithm increased significantly, the correct
rate is also improved, there is a better efficiency and
robustness.
Recently, as the rapidly developing of image
information, SIFT algorithm has made a great
contribution to image automatic mosaic and image
fusion, based on feature matching, we can move
forward apply SIFT algorithm to practical
application, as target recognition and image data, ect.
Provide essential conditions for image analysis and
other technology.
ACKNOWLEDGMENT
The research work is supported by A Project of
Natural Science Foundation of Shandong Province
(ZR2011FQ038), China, and the National Natural
Science Foundation of China (Grant No. 61502259),
and the Project of Shandong Province Higher
Educational Science and Technology Program
(J10LG20, J12LN09), China. Jinyong Cheng is the
corresponding author.
REFERENCES
Bay H, Tuytelaars T, VanGool L,2006. SURF
speeded
up robust features. Proceedings of the 9 European
Conference on Computer Vision.
Chen Y, Shang L,2015. Improved SIFT image registration
algorithm on characteristic statistical distributions
and consistency constraint. Optik - International
Journal for Light and Electron Optics
David G.Lowe. 2015. Adaptive registration algorithm of
color images based on SURF. Measurement.
Glareh Askarieh,My Hedhammar,Kerstin Nordling, etc.
2010. Self-assembly of spider silk proteins is
controlled by a pH-sensitive relay. Nature.
HeMY
DaiYCZhangJ. 2008. Image registration by
integrating similarity and epipolor constraints.
Proceedings of The 3rd IEEE Conference on Industrial
Electronics and Applications.
Huang L, Chen C, Shen H, etal. 2015. Adaptive
registration algorithm of color images based on SURF.
Measurement.
Jiachang Gong, Jichang Guo, 2016. Image copy-move
forgery detection using SURF in opponent color space.
Transactions of Tianjin University
Koiolovoci, Vagvolgyib, 2000. DApplication of
panoramic annular lens for motion analys is
tasks:Surveillance and smoke detection. IEEE
DThornburg Jonathan, Rodes Charles, Lamvik Michael,
etc, 2006. Image analysis method (IAM) for
measurement of particle size distribution and mass
availability on carpet fibers. Aerosol Science and
Technology.
Tony Lindeberg, 1994. Scale-space theory: A basic tool
for analysing structures at different scales. Journal of
Applied Statistics
Automatic Registration Method for Leather Section Image using SIFT and Wavelet Transform
57
Automatic Registration Method for Leather Section Image using SIFT and Wavelet Transform
57