Image Stitching: Research And Application

Image stitching technique is an important research field in computer vision and computer graphics.It has broad applications in the fields of panoramic picture synthesis,video stablization,medical image registration,and virtual reality.Image stitching algorithm is generally divided into registration and blending.The two steps directly affect the final result of image stitching.In recent years,more and more attention is paid to multi-view registration algorithm and multi-plane registration algorithm,for improving the practicality of image registration algorithms.One of the key issue in the image registration algorithm is how to use a more accurate model to align images and reduce registration errors.The translational motion of the camera can cause parallax,and image stitching scenes with varying depths cannot use global saliency for accurate registration.These problems are full of challenges.In the field of image blending,the smooth transition algorithm has a good practical effect.The seam-cutting-based image blending algorithm has received extensive attention in recent years.By the combination of image registration algorithms,these image stitching algorithms provide many new ideas for image registration.Based on a large number of new research achievements in recent years,this paper investigates and collate classical literature and the latest literature in the field.Image stitching algorithms are classified according to image registration and image blending.According to different scene models,the image registration algorithm can be divided into single-view image registration,multi-view image registration and multi-plane image registration.The paper briefly introduces and compares some representative algorithms.Image blending algorithms based on seam-cutting are not limited to improving image blending effects in modern image stitching algorithms.Algorithms that combines seam-cutting and registration can improve local registration performance.In addition,the researchers started the study in terms of avoiding perspective distortion,protecting the natural look of the image,and protecting the geometry in the image.Through the research and summary of the field of image stitching,the paper systematically tease out the technology route of image stitching.For the application of panoramic video stitching,this paper developed a super-wide-format panoramic video stitching and stitching plug-in based on Adobe's nonlinear editing software After Effects with the single-view image registration model.The plug-in possesses non-linear video editing ability with After Effects,supports multiple video formats for input and output,and supports ultra-high resolution video formats.Through the plug-in,up to 12 channels of video can be input and edited at the same time.It also supports frame synchronization,region of interest selection,projection plane selection,blending algorithm setting,and preview and trimming,providing video producers with easy-to-operate panoramic video production solutions.By creating an actual project demonstration,the paper demonstrates the system's production process and show the actual results.For the local image registration,this paper proposes a grid image stitching algorithm with linear structure protection(MISwLP,Mesh-based Image Stitching with Linear Structure Protection).The current image stitching algorithms generally ignore the protection of the important geometric structures in the image,resulting in the phenomenon of geometric distortion in the stitching results,which affects the visual perception.MISwLP is based on the grid deformation stitching algorithm.First,images are divided by mesh.Second,a grid optimization model is established by defining an energy function for the vertex coordinate set of the grid.The energy function consists of local similarity term,global similarity term,alignment term and linear structure term.Finally,the conjugate gradient least-squares method was used to minimize the energy function to get an optimal mesh vertex set to guide the mesh deformation.We conduct image stitching experiments for pictures in different scenarios and compare them with several popular image stitching softwares and algorithms.The results show that,compared with AutoStitch,the image stitching algorithm based on grid optimization can adapt to more complicated multi-planar scenes,and it performs better in reducing projection distortion and alignment error.Compared with SPHP(Shape-Preserving Half-Projective Warps for Image Stitching),APAP(As-Projective-As-Possible Image Stitching with Moving DLT),and NISwGSP(Natural Image Stitching with the Global Similarity Prior),MISwLP algorithm not only aligns images well but also avoids projection distortion,and can maintain the consistency of the image overlapping area to the non-overlapping area,which protects the linear structure in the original image.