Implementation Of Improved SIFT Mosaic Algorithm

With the development of image processing technology,people have been unable to satisfy the image obtained by a single sensor.More and more large-scale scene images generated by using stitching technology are applied to all aspects of our lives,this stitching technology have been widely used in such as medical,automotive electronics,remote sensing aerospace,now more popular virtual reality(VR)and other fields.Image stitching is a technique of spanning a large,seamless and high-resolution image with two or more images of overlapping areas(these pictures may be obtained at different times,different viewing angles,or different sensors).Two key technologies in this process are image registration and image fusion,in which image registration is a prerequisite for image fusion.In order to achieve accurate image fusion,we must first image registration transformation.This thesis summarized the basic flow operations of image mosaic,including image preprocessing,image feature point extraction and matching,and image fusion.SIFT algorithm has never been transcended owing to the advantages of scale-invariant and noise-insensitive and has an extensive application.However,there is a high time complexity for SIFT.In this thesis,the progress of implementing the algorithm was introduced in detail based on SIFT algorithm,and three alternative improvements to enhance the matching speed were proposed.The first point is that,in the initial stage of extracting feature points,this thesis preliminarily determines whether the point is the feature point by comparing the relative size of the center point and the gray value of the pixel points around the circumference of the center.The second point is that in the feature points of the re selection phase,mainly to eliminate the local dense regions of the excessive feature points.In the area of local feature points,one by one to calculate the relative gray level fluctuation of the points around the feature points,that is the gray variance,and the corresponding maximum points of the local area are preserved as the characteristic points of the region.In this paper,we will construct a K-d tree to the image being matched,and then search the feature points in the image by the BBF optimization algorithm.After the RANSAC algorithm is refined,the single stress matrix of the image is obtained.In the phase of image fusion,the third improvement was proposed.Adaptive median filter for image smoothing consists a large degree of retaining the edge information,in addition obviously the effect of noise rejection.It can be summarized that the proposed algorithm has great potential application through simulating on different scenes with different complexity.Note and analyze how the changes in image switching,noise,and illumination effect on the time complexity,and in contrast to the original SIFT algorithm,there is an improvement that reduce the time complexity without effect on image stitching and the robustness of noise.