| Digital image registration is a branch of computer vision that deals with the geometric alignment of a set of images. A large body of research has been drawn to this area due to its importance in remote sensing, medical imaging, computer graphics, and computer vision. In particular, image registration algorithms for applications such as stereopsis, motion estimation, moving object detection, mosaicing, robotics, and vehicle navigation, have received considerable attention. Despite comprehensive research spanning over thirty years, robust techniques to register images in the presence of large deformations remains elusive. Most techniques fail unless the input images are misaligned by moderate deformations.; The primary objective of this thesis is to investigate robust image registration algorithms for geometrically aligning images subjected to large perspective deformations. State-of-the-art techniques are generally limited to small deformations. Although log-polar techniques have been proposed to accommodate rotation and scale, its use in registering images subjected to perspective distortion has not yet been explored. This thesis introduces the use of log-polar techniques to invert perspective deformations among image pairs. We achieve subpixel accuracy through the use of a hybrid algorithm that features log-polar mappings and nonlinear least squares optimization. The registration process yields the eight parameters of the perspective transformation that best aligns the two input images. Together these eight parameters completely characterize all naturally occurring geometric transformations in the real worlds of natural and artificial perception. In this thesis, the addition of general perspective transformations fills the transformational gap left by previous methods which were limited to small deformations.; This thesis contributes several results, including: (1) the introduction of log-polar transforms for inverting large-scale geometric transformations; (2) a study of the effects of affine and perspective transformations in the log-polar domain; (3) a review of the use of various similarity measures, including SSD, cross-correlation, correlation coefficient, and mutual information; (4) investigation of the use of color information to enhance registration accuracy; (5) analysis and comparison of the proposed method against state-of-the-art techniques; (6) an efficient multiresolution implementation; and (7) extensive evaluation over 10,000 image pairs.; The scope of this work shall prove useful for various applications, including the registration of aerial images, the formation of image mosaics, motion stabilization, and diminished reality. |
