Research On Multiple Homography Image Alignment And Eliminating Structure Misalignment In Image Stitching

Image stitching is to combine a few images, it projects images onto a common surface and then fuses all the adjacent images to generate a panorama.The traditional image alignment finds the corresponding pixels between images by SIFT feature matching and estimates homography on those feature points. Taking into account that the SIFT features are mostly detected in large gradient region and a number of points are excluded among transformation matrix estimation, this thesis presents that SIFT features are detected only in larger gradient area to improve processing speed. Gradient threshold is determined according to the number of pixels to keep and the minimum gradient. Transformation between two images is usually represented by a homography. But homography only produces perceptually correct composites when the scene is planar or when the views differ purely by rotation. So we apply multiple homography. We discover multiple homography by interactively applying strict Random Sample Consensus in a divide and conquer manner. After obtaining multiple homography, we divide the image to many rectangle blocks and use the nearest homography.After remapping to a common surface, images need to be fused together. Optimal seam is a common method. We apply dynamic programming algorithm to find the Optimal seam. When the image is not aligned well, there will be disorder or broken geometry in the joints. Pixels in the seam are matched based on local energy and then sparse deformation vectors are derived. The deformation vectors of other pixies in the overlapping region are obtained by Poisson fusion. At last, we do content-preserving deformation to align the disordered structure.