Image Registration Based On Low Rank Subspace Partition

Image registration aims at matching or superimposing images from different sensors,climates,illuminations and viewpoints.In recent years,with the development of computer technology,image registration is applied to many tasks such as medical science,remote sensing data analysis and computer vision.As known to all,algorithms of image registration that are based on interest points have gained popularity and will have wide application prospects.Thus,in this paper,we present a new method based on interest points for image registration.The image registration based on interest points is comprised of interest point extraction,interest point matching and essential matrix estimation.Since there are many problems and drawbacks with existing methods,we try to enhance the extraction and matching of interest points for image registration.Our research is separated into two components as follows:1)Local features have been widely utilized to detect interest points and produce descriptors.Many existing methods detect interest points by local features,and then extract descriptors from local patches around detected points.However,since these methods suffer from sparse noises,such as illumination changes and occlusions,and only employ local features,there are the robustness problem with interest point detection and the indiscrimination problem with patch-based description.As we all know,each image is composed of a clean image and sparse noises and the clean image is in a low-dimension manifold subspace.Besides,it is hard to define interest points in natural images with human annotators.Thus,we present a low-rank deep network(LR-Net)based on the low rank subspace partition.We incorporate a lowrank recovery module to remove sparse noises and get clean information,and then we construct robust global geometric representations efficiently.Finally,we utilize robust local features and global geometric representations to detect and describe interest points.In the training of the model,we use a Siamese network with a selfsupervised training strategy to enable detection and description of interest points to be trained automatically.Experiments on several challenging benchmark datasets show that our learned LR-Net outperforms the state-of-the-art methods.2)Many existing methods match descriptors of interest points to obtain initial interest point correspondences and then find good interest point correspondences that are key for essential matrix estimation.Finding good interest point correspondences requires local and global information of interest point correspondences.However,it is hard to extract local and global information with irregular distribution of correspondences.In this work,we propose an end-to-end correspondences classification network based on local correlation,hereafter referred to as LC-Net,to infer the probabilities of interest point correspondences being inliers.To address the issue of information extraction,this proposed network is built with two novel operations.First,we introduce a novel module,called Local Correlation,to mine strongly connected interest point correspondences for robust local feature extraction.Second,we incorporate a series of feature extraction blocks into the neural network to extract robust local features and global features to find good interest point correspondences for image registration.Our experiments on multiple challenging datasets demonstrate that our method is able to drastically improve the state of the art with little training data.In this work,we finally integrate the above components into a system for image registration and our method outperforms many existing methods that are widely accepted.