Research On Panoramic Image Stitching Algorithm

With the development of technology,computer technology has been widely used in various fields.As a key application field of computer vision,image stitching refers to the process of combining images with overlapping regions to obtain a wide field of view and high-resolution image.So far,the existing image stitching algorithm has a good performance in planar scene while there are certain defects in practical applications of non-planar scenes.And these algorithmic defects may cause issues such as shape distortion and ghosting in the stitched image.And the degree of image stitching distortion is closely related to the scene difference.The larger the scene gap.the more likely the distortion is.Especially for non-planar scene stitching,the distortion problem is particularly serious.In addition,the primary goal of image alignment is to ensure that the overlapping areas are properly aligned to avoid ghosting.The existing processing methods in non-planar scenes are to use local projection transformation to ensure the overall image continuity and to ensure the overlapping regions without misalignment.However,the case where the edge region is stretched and deformed is still difficult to avoid.And in image stitching,most of the algorithms use the global homography matrix to transform the coordinates of the entire image while such methods are only applicable to the case of a planar scene or images obtained only by the rotation.And the spliced images will have obvious ghosting and misalignment when the images don't completely conform to the premise of the transformation model.Therefore,solving only the stitching problem in a planar scene is insufficient to meet the needs of practical applications and how to improve image stitching technology to adapt to more application scenarios has become an urgent problem to be solved in the field.In this thesis,we improve the traditional image stitching algorithm on the limitations of the current algorithm.The contents are mainly in the following three aspects:(1)Aiming at the mismatching problem in traditional feature point pairs extracting algorithms,we propose an image registration algorithm based on the feature matching through topological constraint optimization.And give the experimental comparison results between the improved algorithm and the traditional algorithm.In addition,we utilize the Feature Point Repetition Rate and Recognition Rate to evaluate the registration results.And the experimental data is used to compare the performance of the algorithm with traditional method in feature extraction and feature matching.The results show that the topological constraint splicing algorithm has significant improvement in the feature point matching accuracy compared with the traditional algorithm.The results show that compared with the traditional algorithm,the topological constraint splicing algorithm has a significant improvement in the accuracy of feature point matching.(2)For the problem of linear deformation of image caused by traditional stitching algorithm,we propose an image mosaic method based on the grid linear structure protection and the neighborhood weighted optimization.We first mesh the image and establish an optimization model,and then define the energy function for the mesh vertex coordinate set.The gradient vertices are obtained by the gradient least squares method,which is used to guide the mesh deformation.Finally,the stitched image de-ghosting is performed by neighborhood weighting.(3)For the fact that the panoramic image stitched out by the existing stitching algorithm has irregular boundary,we propose an algorithm for optimizing rectangular panoramic images by utilizing mesh deformation.We first obtain a preliminary rectangular image by local deformation,and the main purpose of which is to facilitate placing the grid on the input image.Then,we use the global deformation optimization mesh to maintain the content characteristics such as shape and line in order to preserve stitching effect of the image while retaining more information.In this thesis,we proposed an image registration algorithm based on topological constraint optimization feature matching,which improves the accuracy of feature point matching.At the same time,the proposed image stitching algorithm based on grid linear structure protection and neighborhood weighted optimization significantly improves the deformation and ghosting of linear structures in images.Finally,we utilize mesh deformation to optimize a perfect rectangular panoramic image,retain more image information,and achieve better stitching effect.