| image registration is a basic research content in the field of computer vision.Registration is to map one image to another image through some kind of spatial transformation,so that the points at the same position in the two images correspond to each other.At present,few researches have been carried out in the field of registration for scenes such as star maps with sparse point sets and cloud maps with many low-texture areas,and there is no targeted method.In view of this,this paper focuses on the registration method based on image features,and has made the following progress:In the field of deep-space star maps,this paper proposes a sparse point set matching algorithm based on Delaunay subdivision,which uses the geometric topological structure between star points to solve the image transformation parameters and achieve preliminary registration.In the detection of star points,multi-scale features are combined,and then the star points are divided into triangular meshes,and finally the star points are screened according to the similarity of the triangles to obtain the result of star point matching.In the process of removing mismatched star points,I learned the idea of iterative nearest point algorithm.While solving the image transformation parameters,iterate continuously to complete the precise registration,and then calculate the image transformation matrix.Experiments show that combining the matching algorithm with the parameter solving algorithm can still have a good registration result in the presence of noise,target points,and missing star points in the star map,which can meet the requirements of registration accuracy.For low-texture images such as cloud images,this paper proposes a new network architecture.On the basis of the Swin-Transformer network,a branch to calculate the image mask information is added,and the contribution of each pixel to the registration is screened to obtain Weighted fusion feature map.Then the homography matrix between the images is calculated and output via the ResNet network.The network can accept image input of any size.The core idea is to allow the network to focus on areas that have a greater contribution to the calculation of the transformation matrix in the cross-window attention map,thereby improving the accuracy of the registration results. |
