IMAGE STITCHING

ABSTRACT:
The process of merging two or more images into a single image is called as Image Stitching. It is often
known as panoramic image. It is done in two ways, in the first approach we have to stitch images
based on intensities of pixels this is popularly known as Direct technique, in the second one we have
find features which define the image and combine the images based on them. It is usually performed
through software of computer.The final image should look like a single image rather than
overlapping of images. There are algorithms which does this process precisely delivering high
definition images. Scale Invariant Feature Transform abbreviates as SIFT is one of them. Image
stitching is an area in computer graphics. The output image is also known as mosaic.In these days
electronic devices are equipped with these features giving wide-angle panoramic images. These
images are used in maps and satellite photos. This paper helps you to get a clear idea on feature
based image stitching.

Key-words:Image Stitching, panorama, SIFT, Direct technique, Mosiac.

1.INTRODUCTION:
In the beginning people used to stitch images manually which consumes more time and does not
give the good quality image, with the advancement of technology we are able to get high quality
images with the help of digital cameras and some mobile phones. Panorama is a Greek originated
word, ‘pan’ means everything and ‘horama’ means view. The perfect overlap of consequtive images
produce great results. Copying and attaching of ovetlapping images may give visible seams due to
some errors, panoramic image stitching provide perfect results by cutting and blending the source
image.

Feature based image stitching focuses on invariant features these are the features which does not
change even if the image transformed, translated, maximized or minimized. This method has high
computational speed and also gives high definition image. This technology is also used in Image
classification also known as image processing, Virtual reality, satellite imaging, Constructing 3D
image and in Medical analysis. In medical analysis they need microscopic details to know each and
every aspect of the patient this technology helps to give complete image with high quality. It is used
in treating patients with the disorder of heart, lungs, liver, pelvic, retina etc.

This process is mainly done in two steps the first one is Image registration in this process we find the
correspondence between two images. In this we draft points between one image to matching points
in other image geometrically, we have to take pictures from various electronic devices from different
angles to obtain all possible perspectives

in 2D and 3D. The second step is Image blending in this process there must be a smooth transition
between two images without any distortion. This is done with the help of alpha blending which
focuses on image pixels or Gaussian pyramid which focuses on Frequency shift.

2. Literature survey

In past few years several researchers presented multiple image stitching processes. Weiss and Lewis
came up with various cost functions for calculating the quality of image stitching. The goal of image
stitching to provide high definition mosaics. The mosaics should be geometrically similar to source
image and the two images should be blended perfectly without any seam. These are not considered
fully in previous approaches. The quality of the image stitched is calculated by the Gradient Domain
in the blended region. Scale Invariant Feature Transform known as SIFT algorithm given by Lowe in
the year 1999. It is based on Invariant Features means features will not change even after
transforming, scaling out or scaling in the image. These are called as key points and the around them
is called as descriptor. The limitation of this algorithm is complexity as it considers 128 dimension
feature vector. This reduces the computational efficiency. To overcome this problem SURF( Speed up
Robust Features ) is invented in the year 2006.

SURF is faster than the above algorithm. It has more computational power and relative robustness to
SIFT. After many developments multiple image blending is developed. Its target is to blend low
frequency to large special area and high frequenncy to small special area . Many algorithms used to
find the seam between the overlapped images. Dynamic programming and many algorithms were
used to reduce the seam i.e blend images perfectly. Many changes and modifications are involved to
reach this stage.

3. Problem statement:

In this project we take an image it is called as source image we have to crop the image into
two parts be it in any ratio, now we have to stitch those two images using SIFT algorithm by
following the stipulated procedure. After implementing we have to compare the final image
to the source image and see the accuracy of images. Our aim is to stitch two images
seamlessly.

4. Process Flow:

First take the images as input and convert them into Grayscale, in Grayscale the image has colors as
multiple intensities of grey based on the original color wavelength. Then find the sift features i.e key
points and descriptors, then find the top few matches. Wrap and stitch the images based on the top
matches. Then display the image by trimming the unnecessary part.

The process flow shows how the whole process of the project is done and the result is viewed. The
5. Methodology:

5.1 Installing the tools and libraries:

Now we need to install the Open Computer vision library (OpenCv) version 3.4.2.16 that is
compatible with the above python version. After Installation we need to import the “cv2” and
“numpy” libraries.These libraries are used to read, write, converting the data into other forms
display and reshape the images in our project.

5.2 SIFT Algorithm

scale invariant feature rework abbreviated as sift is employed to find the keypoints and reckon the
descriptors.keypoint may be a circular image region with associate in nursing orientation. its
represented by a geometrical frame of 4 parameters: the keypoint center coordinates x and y its
scale the radius of the region and its orientation an angle expressed in radians .the sift detector is
invariant or a lot of accurately covariant to translation rotations and re scaling of the image.

A SIFT descriptor may be a 3D spatial bar chart of the image gradients in characterizing the looks of
a keypoint. The gradient at every picture element is thought to be a sample of a three-dimensional
elementary feature vector, shaped by the picture element location and also the gradient orientation.
Samples are weighed by the gradient norm and accumulated in an exceedingly 3D bar chart h, that
forms the SIFT descriptor of the region.
Now these keypoints and descriptors are detected and computed by the built-in perform called
“detectandcompute()”.It returns the list of keypoints and descriptors an

5.2.2 Finding the top Matches:

After the descriptors are known we have a tendency to calculate the space between every descriptor
to all or any different descriptors in image. currently we have a tendency to establish the
correspondences as a result of so as to mix 2 pictures we want to seek out the overlapping
points.These overlapping points can provide US a concept of the orientation of the second image per
initial one. And supported these common points, we have a tendency to get a concept whether or
not the second image is greater or smaller or has it been turned then overlapped, or even scaled
down/up then fitted. All such info is yielded by establishing correspondences.

So for matching we have used BFMatcher methods (Brute Force).

After the matrices are calculated , the algorithm attempts to form a homography matrix between
the two images using four of the matching keypoints. This homography is then used to determine
which keypoint matches are valid and which ones are not based on the error between the projection
of appoint in image 1 with its corresponding point in second image. Using a predetermined
threshold are considered as inliners, and all others as outliners. The number of inliners determines
the quality of the predicted homographic matrix.

5.2.4 Stitching images

Finally we have to stitch the images by trimming the borders. Borders are nothing but the
unnecessary overlapping black region.

6. Results:

We will get the desired image with optimum seaming with image blending smoothly. The output
image will be same as source image.

7. Conclusion:

Image stitching is of creating a high quality panoramic image by amalgamating two or more images.
It is a present research area in the domain of CG. It has various algorithms for feature detection. In
this paper we discussed about stitching of two images by SIFT algorithm, We also discussed about
Image Registration and Image Blending. Scale invariant feature transform algorithm provides
invariant features which does nt change at any cost. We also discussed various steps involves in SIFT
algorithm.