1. UKM Learning and Research Repository
2. PHD Thesis / Tesis PHD
3. Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Title:	A new approach to corner detection and matching for image registration
Authors:	Kahaki Seyed Mostafa Mousavi (P49994)
Supervisor:	Md Jan Nordin, Associate Professor Dr.
Keywords:	Image registration.
Issue Date:	5-Jan-2015
Description:	Corner detection and matching are fundamental steps in many computer vision and image analysis applications. In recent decades, many algorithms have been proposed, with the aim of establishing efficient and robust techniques. The aim of this research study is to provide a robust method for extracting corner points in order to identify the corresponding areas in both the original and the target images. This method can be used in image registration applications, as it will overcome the limitations of the existing approaches. A new image transform, known as mean projection transform (MPT), is proposed in this work, for use as a corner classifier, through analyzing the functions of all vertical and horizontal signals in the local area of the image. In addition, a new nonrotational based projection is proposed, which - unlike other image transforms, such as Radon transform - is not based on image rotation. Instead, it is general and thus applicable to all capturing signal functions. In other words, the signal acquisition is based on projector movement and the capturing function is not limited to integral or medians, as is the case in Radon and Trace transforms. Moreover, only the active parts of the signal that carries information of greater importance are considered in signal acquisition. Then, a linear approximation of the parabolic fit is calculated to localize the corner points of the identified candidates. A deformation invariant feature correspondence method is proposed as a corner matching algorithm. Deformation problems, such as affine and homography, affect the local information within the image and thus result in ambiguous local information pertaining to image points. In addition, existing similarity measurement techniques, such as normalized cross correlation (NCC), squared sum of intensity differences (SSD) and correlation coefficient (CC), are insufficient for achieving adequate results under different imaging conditions. Thus, new descriptor's similarity metrics, which are robust under local variations and deformations, are proposed to establish an efficient feature correspondence over multiple images. Firstly, a normalized Eigenvector correlation (NEC) based on the Eigenvector properties of the signals is developed. Then, a signal directional differences (SDD) based on processing the signal direction changes is introduced. Finally, to improve the mutual information-based techniques, a rotation invariant mutual information (RIMI) based on information theory is proposed. Eventually, a new framework for image registration based on the aforementioned techniques and correspondence approval is developed. The proposed corner detection, corner matching and image registration application are tested on standard test images sourced from 'Featurespace' and 'USC-SIPI' benchmark datasets. The results indicate that the new methodology produces fewer false-positive and false-negative points compared to the state of the art techniques. Moreover, a new corner detection evaluation metric called accuracy of repeatability (AR) is introduced. AR combines the repeatability and localization error for finding the probability of each detected corner point in the target image. The output results demonstrate that the proposed techniques outperform the other methods tested on standard datasets using repeatability, localization, AR, precision-recall, corner correspondence and image quality assessment techniques, such as PSNR and SSIM metrics. Different approaches using different standard datasets and evaluation metrics are used to evaluate the proposed techniques. In order to demonstrate that the method developed as a part of this study is a significant improvement over the existing methods, statistical tests, including T-test and Friedman test, are employed. The findings of the statistical analyses further confirm the robustness of the proposed methods, which is a significant improvement over the other standard techniques.,Ph.D
Pages:	266
Call Number:	TA1637 .K337 2015 3
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.