Aerial image alignment and stitching are among the oldest and most important applications of machine vision. Image alignment is used in every video camera with image stabilization property. Stitching algorithms are used in constructing digital maps and stitching satellite pictures. In constructing panoramas which recently has found important applications in virtual environment, advertising and so on, a combination of these algorithms is used. Meanwhile, each particular application is defined under special conditio while there are efforts to manage these applications automatically and without any preconditions. In early attempts for automatic image alignment and stitching these tasks have been based on the pixel intensity or color. Hence every pixel of images is processed and then the transformation between images is derived. Because of a huge number of variables (pixels) and blind search on the transformation parameters, these algorithms have high order of complexity which leads to expensive memory and time cost. In addition since these algorithms rely on the pixel intensity they are sensitive to illumination, view, rotation and zooming. More recent approaches implement object recognition methods that eliminate the need for complete process of the entire images. Image alignment is achieved by focusing on interest regions, instead. This leads to optimization of time and memory. In addition the methods are insensitive to illumination change and any other transformation that may happen to the images. In this thesis we introduce image alignment and stitching and review some existing algorithms. We describe how object recognition approach can be applied in image alignment. Furthermore we suggest two new algorithms for image alignment and stitching that improve some stages of the algorithms and evaluate them. Key word : Aerial image, image stitching, object recognition, alignment
