Scribd
Upload
132 views

42.image Stitching Using Machine Learning

The document proposes a method to automatically cluster images into separate panoramas based on scale-invariant feature matching (SIFT) without requiring manual separation. It uses SIFT to robustly match images regardless of factors like zoom, rotation, and illumination. Keypoint matching, RANSAC, and clustering are used to discover relationships between matching images and break the image database into individual panorama datasets, which are then stitched together linearly.

Uploaded by

nikil chinna
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
132 views

42.image Stitching Using Machine Learning

The document proposes a method to automatically cluster images into separate panoramas based on scale-invariant feature matching (SIFT) without requiring manual separation. It uses SIFT to robustly match images regardless of factors like zoom, rotation, and illumination. Keypoint matching, RANSAC, and clustering are used to discover relationships between matching images and break the image database into individual panorama datasets, which are then stitched together linearly.

Uploaded by

nikil chinna
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 4

Abstract:

Creation of panoramas using computer vision is not a new idea, however,


most algorithms are focused on creating a panorama using all of the images
in a directory. The method which will be explained in detail takes this
approach one step further by not requiring the images in each panorama be
separated out manually. Instead, it clusters a set of pictures into separate
panoramas based on scale invariant feature matching. Then uses these
separate clusters to stitch together panoramic images.

Existing System:

Algorithms which have been used to create panoramas have been developed
in the past and implemented many times. Some of these algorithms have
even been optimized to run in near real time applications on mobile
processors. Additionally, there are some well-known machine learning
concepts which have been used to classify data into separate groups based
on common characteristics.

There are a number of different classification algorithms which are well


known and widely implemented to solve a number of different problems. In
particular many of these algorithms come from machine learning techniques.
Some examples include support vector machines, and k-means clustering.
These algorithms all need to be tweaked for the given task.
Proposed System:

The current method proposed by this paper uses invariant feature based
approach, SIFT, to automate image sorting and panorama stitching. The use
of SIFT allows robust matching of pictures in the image database
irrespective of camera zoom, rotation, and illumination. This method
discovers the relationship between the matching images so it allows the
image database to be broken down into individual panorama datasets. The
current method takes in an image database and sorts the images based on
relationships using SIFT, keypoint matching, RANSAC, and clustering. Then
the panorama image datasets are stitched together linearly.

SYSTEM REQUIREMENTS

HARDWARE REQUIREMENTS:

 Processor : Intel i3 and above


 RAM : 4GB and Higher
 Hard Disk : 500GB: Minimum

SOFTWARE REQUIREMENTS:

 Programming Language / Platform : Python


 IDE : pycharm/jupyter
Modules:

Preparing Image Database

in this module users can upload the images of their sites which will be stored
in database

Homography Calculations
After the matches are calculated, the algorithm attempts to form a
homography matrix between the two images using 4 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 a point in image 1 with its corresponding point in image 2.
Using a predetermined threshold, all matches which are acceptable based on
that threshold are considered inliers, and all others are outliers. The number
of inliers determines the quality of the predicted homography matrix.

Clustering algorithm
In order to determine which pictures corresponded to separate panorama’s,
a method for identifying similar images needed to be used. First, each pair of
images is checked to determine the percent of inliers. While the homography
calculation determined which points created the best homography, it didn’t
filter out images which had no good matches, or had a single good
homography by chance. Step one, was to look for images which were likely
to match by checking the percent of the keypoint matches which were
inliers. Any images where more than 50% of the matches were inliers, were
considered candidates for stitching together, and marked as such.
Panorama Stitching
After a directory of all the images in a single panorama has been created,
the two images with the largest number of keypoint matches inside this
directory are used for stitching. The first step in this process is to calculate
the best homography in a similar way to how it was being done before.
Using this homography, one image is warped to be in the same frame as the
other and a new image of all black pixels is created which can fit both
images in the new frame.

You might also like