Image registration

• Main Author: Pua, Denny Zaiqi
• Other Authors: Chua Chin Seng
• Format: Final Year Project
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering
• Online Access: http://hdl.handle.net/10356/61091
• Institution: Nanyang Technological University

To help end users to continue using their small, compact digital cameras, yet still being able to capture the beautiful scene, image stitching is required. Users will only need to capture multiple different parts of the scenario, process the information and images into a program, which will be MATLAB in this project. It will proceed to stitch the images, producing the final image, which will be similar to the beautiful scene that they had initially wanted to captured, based on the images they have input for process. There are many reports and research done over the years regarding image stitching. After extracting the edges and points, computation algorithm has been performed for further use of the points and edges. To extract such points, edge extraction method such as Harris Detection[1], Scale Invariant Feature Transform, or SIFT[2] has been applied to compute the necessary calculation. These edges and interest points are very crucial in image stitching as they are used as common point between the images. Stable key point detection are desired in this methodology, which DOG, D(x,y,σ) is able to produce by differentiating two of the nearby scale-space and convolute it with the image. The key to this formula lies in the k which is a constant multiplicative factor. Once the extremas have been appointed, the actual key points can be obtained through the Local Extrema Detection proposed by Lowe.[2] Other than scale invariant, SIFT is also orientation invariant which the key points generated are not affected by the rotation and scaling.[2, 3] This is due to orientation assigned to each key points that are calculated by the histogram and algorithm. Lowe suggested that the best result is achieved by using 4 by 4 array of histograms with only 8 orientations bin for each array, which derive the 128 descriptors for each key points.[2]