Image Registration Algorithm And Applications Using Image Depth Information

Image registration refers to matching two or more image frames taken at different times, from different viewpoints, or even different sensors. It focuses on the process of aligning images and establishing the relationship of them on both geometric side and intensity side(normally the intensity transformation is not fully necessary, so the image registration process only cares about the position mapping relation of image frames). Nowadays, image registration becomes one of the hot topics in computer vision and gains wide attention on its applications, such as 3D movie, MRI image analysis and 3D reconstruction software in architecture.This dissertation presents research on image registration algorithm based on depth information, leading to better performance in its further applications. The main work will are as follows.Firstly, the related theory, algorithms and implementations of depth estimation utilizing various clues are analyzed, followed by the summarization of them. Moreover, the depth estimation methods based on defocus blurring and machine learning are studied exhaustively. In the former algorithm, we redesigned the energy function and achieved great improvement on the depth continuity property of estimated depth images.Secondly, conventional feature-based image registration algorithms, such as the one based on Scale-Invariant Feature Transformation(SIFT), can produce lots of mismatches. Existed algorithms using RANSAC(RANdom SAmple Consensus) in mismatch removal can achieve better result, while the threshold of cost function remains challenge. Since most of the existed feature matching algorithms are not so powerful and efficient in mismatch removing, a mismatch removal algorithm was proposed which adopted the depth information and local depth continuity property to improve the performance.Thirdly, two common image registration applications are introduced: image super resolution and 3D reconstruction. On super resolution side, the traditional image degrading model is introduced, followed by analysis of some common super resolution algorithms which are divided into 2 classes. On 3D reconstruction side, a study is carried out on the three types of 3D reconstruction algorithms. The experimental results demonstrate that the proposed image registration algorithm outperforms conventional ones on both registration stage and applications.