Research On Image Stitching Algorithm

Image stitching is an important branch of computer graphics that combines a series of images with overlapping regions into a seamless and large view image.It can solve the problem of camera which is unable to adapt to the large view and high-resolution scene because of the restriction of hardware condition.Therefore,it has been extensively in national defense and military services,security monitoring,medical imaging and other fields.This thesis improves the problems in image registration and fusion algorithm,which are core technologies of image stitching.The main research work is as follows:(1)An improved RANSAC algorithm based on determinant point process(DPP)is proposed,which is applied to image registration by exploring the sampling characteristics of DPP.The DPP is able to perform conditional constraints on the initial points selection of the RANSAC algorithm so that the selected points can be distributed uniformly and globally.Since there is no iteration upper limit in the RANSAC algorithm,hence,K-DPP is employed to automatically judge the iterations for saving computing resources.The improved RANSAC algorithm is remarkably effective in eliminating matching features and shows effective robustness in matching images with large parallax.(2)Based on the As-Projective-As-Possible image stitching with moving DLT,a multiple homographic matrix feature matching and detecting algorithm is proposed in this thesis.The proposed algorithm mostly extracts correct matching points between two images to be stitched and performs image registration by exploiting multiple local homographic coordinate transformation models in overlapping region,which guarantees close registration within local parts,and reduces vagueness and malposition.In addition,favorable robustness in image matching is illustrated in large parallax combining the improved RANSAC algorithm proposed in this thesis.(3)An improved nonlinear image fusion method is proposed in this thesis.The simple image stitching is classified at the first stage.Then,the relation of vertex coordinate positions of warped images is employed to determine types of image stitching.Finally,an improved nonlinear image fusion algorithm is designed.The above-mentioned method can rapidly determine the type of image stitching and eliminate stitching gaps at four boundary directions of overlapping field to realize smooth brightness transition in overlapping field.