Research Of SIFT-based Image Stitching And Feature Detection

As a branch field of the digital image process,image stitching has become one of the most crucial research tasks in this discipline. Especially during these 2 or 3decades, the technology about image stitching, which became more advanced with the development of the computer vision technology and computer graphics, has widely used in a various of different fields, for example, military, transportation, kinematic analysis, medical image analysis, geological prospecting, virtual reality technology and remote sensing technology, and etc. Image stitching aims at stitching pictures with similar background and neighboring images to form a new one of full scope,producing an effect previously given by the panorama camera only.Generally speaking, mainstream image stitching contains two different methods:feature extraction based or suture lines based. In the recent several years, researchers largely focus their research hot point on features’ finding, extracting and matching because of the development of the methods about image stitching and fusion. The SIFT(Scale invariant feature transform)-based detection method are widely used because of its great efficiency and high speed. However, this method has no advantage on real-time because its feature parameters are very complex.In order to solve this problem, this thesis introduces several familiar methods about feature detection and simulates them. The reasons causing low real-time ability of SIFT is analyzed, conclusion is obtained that the majority of run time about a SIFT-based image stitching is taken up by the SIFT-based feature detection, then improvement of the SIFT algorithm is made which is based on this conclusion. The improvement is to reduce the sample filters to some extent in the hierarchy of Gaussian and Laplacian image series, i.e. to cut down the number of layers and grades of the image series. The method is verified with differences of images in this paper. It is demonstrated by experiments that the quality of resultant images are guaranteed with proper reduced SIFT-feature points extracted by our new method, which can improve the real-time ability to some extent.Meanwhile, this thesis improves the method of Laplacian Pyramids-based image stitching to some extent. In the original method the suture lines are sought in all layers,while in our new method the suture lines obtained in the upper layer provide the domain of searching in the following layers so that the region of suture lines seekingis reduced. It not only saves the run time, but also gets better suture lines.In addition, this thesis improves the Poisson image stitching with the idea to process differently according to different suture lines situations. Those images which have obvious global suture lines can be optimized by using Poisson image fusion,which can eliminate the trace to large extent. Those images which own local foreign objects can be processed with masks on the area of foreign objects in the traditional suture lines seeking, which can make the suture lines avoid the foreign objects thus produce no traces of going through the objects. This method improves quality of the resultant image.