Research On Image Stitching Based On Dynamic Mesh Optimization And Combined Transformation

Image Stitching is an active research area in computer vision and digital image processing.The goal of image stitching is concerned with seamlessly stitching multiple overlapping images,even in the presence of parallax,lens distortion,and illumination changes.However,the current mainstream image stitching algorithms still have many shortcomings.The artifacts and distortion problems in the parallax image stitching will seriously affect the visual quality of the image stitching result.Although the mesh optimization algorithm can improve the image alignment effect and reduce the artifacts,the local projection transformation no longer has line invariance.Therefore,the result of image stitching will have local distortion when the parameters of mesh optimization are not suitable.Besides,the local projective transformation is a slight offset from the global projection transformation,and it cannot solve the problem of global shape distortion caused by projection distortion.This thesis proposes a novel image stitching algorithm based on dynamic mesh optimization and combined transformation to solve the shape distort problem of parallax image stitching.The algorithm effectively protects the shape and ensures accurate registration,which significantly improves the quality of image stitching results.The main contents of this thesis are as follows:(1)Review of the image stitching algorithms and discuss the problems of current mainstream image stitching algorithms.This thesis focuses on the shape distortion problem of image stitching.Based on the analysis of the existing algorithms of image stitching,mesh optimization and combined transformation algorithms are selected as the basic framework of this thesis.(2)In order to solve the local shape distortion problem of the mesh optimization algorithm,this thesis dynamically solves the optimal parameters of the mesh optimization algorithm when the image alignment error and the line distortion are minimal.Furthermore,the keypoint compensation method is employed to efficiently correct the local shape distortion of the line region.Therefore,the local shape is protected while retaining the registration accuracy of the mesh optimization algorithm.(3)In order to solve the global shape distortion caused by the non-linear scaling of the projective transformation,this thesis analyzes the principle of the geometric transformation and discusses the reason for the non-linear scaling.We find that using a similar transformation instead of projective transformation can effectively protect the overall shape of the image.It is necessary to construct a weighted combination function of global similarity transformation and local projective transformation to protect the overall shape of the image stitching result.This thesis uses the projective distortion direction to determine a reasonable weight for each mesh,hence ensuring the stitching result changes smoothly.(4)The performance of the algorithm proposed is compared with the global projective transformation algorithm,the APAP algorithm and the SPHP algorithm.The subjective and objective evaluation results show that the algorithm proposed can effectively protect the shape of the stitching result and the visual quality more natural.Besides,the image alignment error RMSE is smaller,and the structural similarity SSIM has been significantly improved.