Optimal Seam Based Image Stitching With Large Parallax

Image stitching aims to generate a seamless high-resolution panoramic image with a larger field of view free of distortions or artifacts as fast as possible.It is an interdisciplinary research hotspot,which has important research significance and practical value in the fields of computer vision,geomatics,and computer graphics.However,in large parallax scenes,existing image stitching methods are often accompanied by problems such as misalignment and ghosting.This paper studies a novel and effective large-parallax image stitching method to deal with these issues.We divide the method into three stages: image registration,seam cutting,and image blending.The main contributions are as follows:Firstly,we propose an image registration method based on a hybrid transformation model for preliminary alignment of input images.Specifically,we combine the linearized local homography transformation and the global similarity transformation with the smallest rotation angle linearly to obtain the hybrid transformation model,ensuring that the overlapping areas are aligned while the distortion of the non-overlapping areas are eliminated.Secondly,we propose a superpixel-based seam cutting method,which is used to solve the problems of misalignment and ghosting after image registration.In this paper,a new energy optimization function is designed.The function contains a difference cost that measures the similarity of images.It integrates image information such as color difference,gradient difference,and texture complexity.Subsequently,the problem is solved by mincut algorithm,seeking an optimal seam that passes through high-similarity area,or going along salient objects,or going through specific low texture complexity areas.Finally,we propose an image color blending method to eliminate visible seams caused by color differences,achieving seamless image stitching.We design a weighted color difference propagation strategy based on superpixels,which can transfer the color difference from one image to another and realize natural color transition.In order to verify the effectiveness of the proposed method,we conduct a series of experiments.For image registration,we analyze the hybrid transformation model and compare with other registration methods.In the seam cutting period,experiments such as comparisons with other seam cutting methods,difference cost ablation analysis,superpixel performance evaluation,and parameter analysis are conducted.To confirm the effect of image blending,we compare it with common blending methods and evaluate superpixel performance.The experimental results demonstrate the superiority of our method.In addition,we extend the twoimage stitching to multiple image stitching.