Image Registration Based On Deep Feature And Application In Locating Targets

Image registration is the process of aligning and transforming multiple images into the same coordinate system through an optimal geometric transformation.As an important step in image processing,image registration is widely used in image stitching and fusion,object recognition,object location,and environmental change detection.With the advancement of Internet technology,especially the development of machine vision,some airborne mobile platforms need to locate targets during the navigation process.Positioning and navigation that only relies on satellite signals is easily affected by factors such as signal interruption and radiation interference,resulting in failure of positioning functions.Therefore,the use of visual information to assist in locating targets,with the help of image registration technology,has shown outstanding advantages.In view of the problems of low registration success rate and narrow universality in traditional image registration methods,in order to improve the accuracy and scope of application of image registration,this paper uses convolutional neural network to extract deep convolutional feature of images,and two end-to-end remote sensing image registration algorithms are proposed.Finally,the application of image registration technology in locating targets is outlined,and the application process of the proposed algorithm in locating targets is given.The main contents are as follows:Firstly,a remote sensing image registration algorithm combining local convolutional feature and self-similar feature is proposed,which utilizes spatial context and local location information to filter interference information such as noise and background clutter in image feature.In the feature matching section,a bidirectional correlation layer is used for bidirectional matching,which increases the utilization of matching information while maintaining the consistency of matching directions.Combined with the parameter regression network,the bidirectional affine transformation parameters are estimated,and finally the synthesized parameters are used to achieve image registration.In addition,a remote sensing image registration algorithm based on parameter regression of attention distribution is proposed.Specifically,a pre-trained backbone network and a global context fusion module are used to increase the attention to important areas to obtain salient feature.Secondly,aiming at the problem of wrong matching in the process of feature matching,the two-way matching map after feature cross-correlation matching is combined with the mutual nearest neighbor principle to filter wrong matching and improve the reliability of feature matching.Then,a parameter regression network of attention distribution is proposed to further filter the interference information in the matching feature,and estimate the bidirectional matching map,so as to obtain two affine transformation parameters,and then use the parameters of the average synthesis to transform the source image to achieve the final registration.Comparing the algorithms proposed in this paper with traditional and deep learning registration algorithms on different datasets,the qualitative and quantitative experimental results show that the two proposed algorithms have certain advantages in the application range and registration accuracy for different types of images.