International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169

Volume: 5 Issue: 4 351 354

_______________________________________________________________________________________________

Image Stitching Based on Corner Detection

Poonam D.Mind Prof. P. P. Gumaste

[email protected], [email protected],
(+918308539876) (+919975563868)

Abstract: An image stitching is a method of combining multiple images which are overlapping images of the same scene into a larger image.
Mostly used methods are Harris corner detection method and SIFTS (Scale Invariant Feature Transform) method. In this paper, a study of Harris
corner detection algorithm and SIFT algorithm is done by comparatively in image stitching using similarity matrix matching scheme. Total 30
pairs of different images have been used for their simulation and comparison. The algorithms have been compared with more number of corners
detected in images, number of matching pairs and number of matching time. From the results of simulation it has been observed that SIFT corner
detection method is most efficient in image stitching.

Keywords: Image stitching, Corner detection, Harris corner, SIFT

__________________________________________________*****_________________________________________________

domain are in general faster and can handle well small

I. INTRODUCTION translation, rotation, and scaling. Unfortunately, the
Image stitching is the sub branch of the computer vision. It performance of frequency domain-based algorithms will be
combines two or more different types of images to form one degraded when dealing with scenarios where smaller
single image that is called as panoramic image. The word overlapping regions exist.
panorama is derived from the Greek words pan and Feature-based algorithms contains a small number of
horama. Pan means everything part of image and invariant points, lines, or edges to align images. One
horama means to view that part, and thus it means all significant advantage of these algorithms is that the
round view. Panorama images can be created in a many computational complexity will be decreased due to less
ways, from the first round painting in the 18th and 19th information that needs to be processed. Additionally,
centuries. The aim of stitching of images is to increase feature-based algorithms are robust to changes intensity of
resolution of image as well as the field of view; people used the image.
image stitching technology in topographic mapping. A Here, Firstly detect corner of the input images, then
topographic map is a type of map which is characterized by perform stitching of image by matching the corners.
large-scale detail and quantitative representation of relief,
using contour lines. Typically, a camera is able to taking II. LITERATURE SURVEY
pictures within the scope of its view only; it cannot take a In the last few years many researchers have implemented
large picture with all the details of image fitted in the one and proposed various panorama image stitching systems.
single frame. Chia Chen [4] has given comparisons of
Panoramic image resolves this problem by combining different new techniques of Image Stitching. Along with
images which are taken from different sources into a single image stitching, the paper also describes methods which can
image. Such type of images are useful for surveillance be used for image mosaicing. Different acquisition
applications, remote sensing etc. Image stitching algorithms techniques of images such as image acquisition by rotations
create the high resolution photo mosaics used to produce of the camera, image acquisition by camera translations, and
digital maps and satellite photos. Creating the high image acquisition due to a hand held camera and their
resolution images by combining smaller parts of images are properties in detail are discussed. The mathematics of image
popular since the beginning of the photography. registration is discussed in detail in the paper. Different
To stitch images and form a panoramic image, the methods of the image registration using different similarity
similar part of the overlapping portion among adjacent measures such as images registration using sum of squared
images needs to be calculated in the first place. Intensity- differences, image registration using sum of product, image
based algorithms usually contains a large amount of registration using standard deviation of differences, by
computation and therefore they are not appropriate for restricting the search set, image registration with step search
image alignment when there is image in rotation motion and strategy have been discussed in brief. D.L.Milgram have
scaling. On the other hand, algorithms based on frequency- been given the image merging method of differences of

351
IJRITCC | April 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169
Volume: 5 Issue: 4 351 354
_______________________________________________________________________________________________
linear distribution of intensity which uses a linear ramp to blending process was done using Discrete Wavelet
spread the intensity differences of the pixels which are Transform (DWT).
immediately next to the seam for blending pairs of grey Antony and Surendran [13] implemented a stitching
level satellite. The paper gives an idea that the concept of technique to create panoramas of satellite images which are
image stitching or panorama production can be used. based on image registration. They geometrically aligned one
Pranoti Kale et al. [10] has given a analysis of image input image into other image. Then, image stitching
stitching algorithm. The paper gives a brief review on the algorithm takes the alignment estimation that is produced by
image registration techniques used in the past as well as in the registration algorithm to blend the images in a seamless
the present area. It has been discussed that the image manner. Their image stitching system was well suited for all
stitching process can be divided into three main steps of types of images including the satellite images. The system
image calibration, image registration and image blending. supported images of different formats, such as JPEG, TIFF,
The main approaches involved in image stitching viz. direct GIFF, and PNG.
and feature based techniques are discussed. The direct Suen et al. [14] show that how the curvature values can
techniques work by directly minimizing pixel to pixel eliminate the effect of non-uniform inconsistency. They
dissimilarities. generated a method that is minimized the curvature value
And also they are not invariant to image scale and variations between the input images and the mosaicing
rotation. Whereas the advantage of direct techniques is that image. The experiment showed that it could reduce
they make the use of the information available in the image conspicuous cutting curves. Moreover, even when there is
alignment. The feature based techniques work by extracting severe geometric misalignment, by choosing an optimal cut
a sparse set of features and then matching them to each between the input images, the induced artifacts become
other. This technique is applied by establishing invisible. In addition, their methods provide an easy control
correspondences between points, lines, edges, corners or of fidelity and transition smoothness by simply determining
other geometric entities. The techniques namely Harris the area of using the minimization.
corner detector, SIFT, SURF, FAST, PCA-SIFT and ORB
come under this technique. III. METHODOLOGY:
Uyttendaele et al. [9] presented two main For stitching two images, need to detect the corners of
contributions of image stitching problems. The first one is a each image. For corner detection we are using Harris corner
method for dealing with objects that move between different detector and SIFT which is explained below :
views of a dynamic scene. The other is a method to 1. Harris corner detector:
eliminate visible shifts in brightness. They presented a Harris corner detection algorithm was proposed by
method of block-based adjustment, which changes the pixel Harris C and Stephens MJ in the year 1988. It is an
values using a weighted average of lookup tables from close algorithm based on still image used for combined corner and
parts of the image. edge detector. Reasonable amount of corner features are
Rankov et al. [10] proposed an approach for extracted which gives a better quantitative measurement by
establishing high resolution, edgeless, and composite image using a stable operator. A local detecting window in image
using cross-correlation and blending. One image is is designed.
correlated at a time with a composite image. The blending is The average variation in intensity is determined by
performed, when the image is registered. The presented shifting the window by a small amount in different direction.
method is fast because of using a lookup table technique. The centre point of the window is extracted as corner point.
Zomet and Peleg [11] studied the cost functions and The point can be recognized easily by looking at the
compared their performance to different scenarios both intensity values within a small window. Shifting the window
theoretically and practically. Their approach can be used in in any direction gives a large change in appearance. Harris
many applications, such as building the panoramic images, corner detector is used for detecting corners.
object blending, and removing of compression artifacts. On shifting the window if its a flat region than it will
Bind et al. [12] proposed a panoramic image show no change of intensity in all direction. If an edge
stitching technique for three-dimensional, rotational images region is found than it will show no change of intensity
with a variation of the illumination. The input overlapping along the edge direction. But if a corner is found than there
images are passed through two strong stitching algorithms, will be a significant change of intensity in all direct ions.
that are SIFT and SURF. SIFT algorithm is invariant Harris corner detector gives a mathematical approach for
towards scale and rotational variation. It is also robust determining the region is flat, edge or corner. Harris corner
towards the noisy environment. SURF algorithm has very technique detects more features and it is rotational invariant
similar properties as SIFT. It has the properties of and scale variant.
illumination invariance and good computational speed. The
352
IJRITCC | April 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169
Volume: 5 Issue: 4 351 354
_______________________________________________________________________________________________
To extract corner can give prominence to the important I2. SURF (Speeded Up Robust Features):
information. Those can be described by equation below. As for the different methods of stitching, image
registration fall broadly into three methods: gray
(, ) = (, )[ + , + , ] information based, transform domain based and feature
based. Among them, the feature-based image stitching
.....(1) technology is widely used because it has the quality of high
...... time efficiency, maximum matching accuracy
Where, and good robustness.
E is the difference between the original and the moved Point feature is an important feature of the image in a
window. various image features; it has the benefits of rotational
u is the windows displacement in the x direction . invariance, not varying with changes in light conditions and
high speed.
v is the windows displacement in the y direction . The common feature points are Harris corner
w(x, y) is the window at position (x, y). This acts like a detection, SIFT (Scale- Invariant Feature Transform) and
mask. Ensuring that only the desired window is used. SURF (Speeded Up Robust Features). For all above features
I is the intensity of the image at a position (x, y). has compared the commonly used local invariant features
I(x+ u, y+ v) is the intensity of the moved window. and found SURF feature detection is more effective than
I(x, y) is the intensity of the original. other feature detection. So in this project, we propose a fast
Weve looking for windows that produce a large E value. stitching method based on SURF. It can be majorly divided
To do that, need to high values of the terms inside the into four steps: feature points extraction, feature points
square brackets. expand this term using the Taylor series. matching, determining the transformation relationship and
image fusion.
(, ) =,[ I(, ) + + ( , )]
...........(2) 3. SIFT descriptor :
SIFT was first presented by David G Lowe in 1999. SIFT
tucked up this equation into matrix form algorithm is very invariant and robust for feature matching
with scaling, rotation, or affine transformation.

, = [ ]
We utilize SIFT feature points to find correspondent points
of two sequence images. The SIFT algorithm is described
...........(3) through these main steps:
scale space extrema detection, accurate key point
localization.
After that 1) Scale space extrema detection:
rename the summed-matrix, and put it to be M: First, we build the pyramid of image by continuous
smooth with Gaussian mask. DoG (Difference of Gaussian)
= (, ) pyramid of the image will be obtained by subtraction

adjacent smoothed images. By comparing each pixel of
...(4) current scale with upper and lower scales in the region 3 x 3,
i.e. 26 pixels, find the maximum or minimum value among
Harris corner can be defined as the maximum in local area them. These points are also considered as candidate of key
by the following formula: point. The equations below will be used to describe
Gaussian function, scale space and DoG.
R = Det(M) k Trace (M) .....(5)
1
, , = 2+ 2) ....(8)
2 2 (
Where,
Det (M) = .......(6) , , = , , (, )....(9)

Trace M = + .......(7) Where * is the convolution operation in x and y.

353
IJRITCC | April 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169
Volume: 5 Issue: 4 351 354
_______________________________________________________________________________________________
, , = , , (, , ) (, ) So if inequality (15) fails, the key point is removed from the
= , , (, , ) candidate list.
......(10)
2.Accurate key point localization: IV. CONCLUSION
The initial result of this algorithm, considers key point In this research work we have performed image
location is at the central of sample point. However this is not stitching using two corner detection method namely Harris
the correct maximum location of key point then, needed a corner detector and SIFT descriptor.
3D quadratic function to fit the local sample points to The image stitcher provides a cost effective and
determine the true location, i.e. sub-pixel accuracy level of flexible alternative to acquire panoramic images using a
maximum value. panoramic camera. The panoramic images stitched by a
Taylor expansion of the scale space function is shifted so the stitcher can also be used in applications where the camera is
original is at the sample point. unable to obtain a full view of the object of interest. The full
view of the object can be constructed using the image
1 stitched using overlapping regional images acquired for the
=+ + /
2 2
object.
Based on extracting invariant scale features, get
.....(11)
potential feature matches for SIFT algorithm than that for
Where, D and its derivatives are evaluated at the sample
Harris algorithm. SIFT can give better performance and
point and x = (x, y, ) T is the offset from this point.
when there are less rotations Harris corner detection
The location of the extremum, x, is determined by taking the
algorithms can perform better.
derivative of this function with respect to x and setting it to
zero, giving
REFERENCES
[1] Donggyu Sim, Yongmin Kim, Detection and
x = D / x D/ x compression of moving objects based on new panoramic
image modeling, Image and Vision Computing, Vol.27,
.........(12) pp. 15271539, 2009
[2] Jiaxi Wang and Junzo Watada, Member, Panaromic
The next stage attempts to eliminate some unstable Image Mosaic Based On SURF Algorithm 978-1-4799-
points from the candidate list of key points by finding those 7253 1/15/$31.00 2015 IEEE
that have low contrast or are poorly localized on an edge. [3] Minchen Zhu, Weizhi Wang, Binghan Liu, and Jingshan
Huang, A Fast Image Stitching Algorithm via Multiple-
For low contrast point finding, we evaluate D(x) value with
Const raint Corner Matching, Hindawi Publishing
threshold. By substituting two equations above, we have: x Corporation Mathematical Problems in Engineering, vol.
1
x = D + D / x x 2013, pp. 1-6, sep 2013.
2 [4] Pranoti Kale, K.R.Singh A Technical Analysis Of
....(13) Image Stitching Based On Different Corner Detection

If the value of D(x) is below a threshold, this point will Methods Vol. 3, Issue 4, April 2015
be excluded. To eliminate poorly localized extrema, use the [5] Tejasha Patil, Shweta Mishra ,Poorva Chaudhari ,
fact that in these cases there is a large principle curvature Shalaka Khandale Image Stitching Using Matlab
across the edge, but a small curvature in the perpendicular International Journal of Engineering Trends and
direction in the difference of Gaussian function. Technology- Volume4 Issue3- 2013
A 2x2 Hessian matrix, H, computed at the location and scale
of the key point is used to find the curvature. With these
formulas, the ratio of principle curvature can be checked
efficiently.

=

.... (14)

( + ) /Dxx Dyy (Dxy) < (r + 1)/r

......(15)

354
IJRITCC | April 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________