Automatic Registration Of Multimodal Image Using Geometric Structural Properties

The advancement of sensor technology has promoted the development of remote sensing technology,the number of remote sensing images of various modalities is increasing.At the same time,due to the limitation of its single imaging band,single-mode imaging is increasingly difficult to meet daily needs.Therefore,it is necessary to make comprehensive use of multimodal remote sensing images to complement each other’s advantages.Multimodal image registration technology is the basis for the comprehensive utilization of multimodal remote sensing images,and it is also a hot research topic in the field of remote sensing.However,different wavebands have different spectral characteristics and radiation characteristics,so there are significant nonlinear intensity differences between multimodal images.Moreover,the sensor imaging mechanism makes the presence of noise in the multimodal image inevitably,resulting in image degradation.These problems make the following difficulties in multimodal image registration: high repetition rate feature extraction,modal invariant descriptor construction,robust similarity criterion and high-precision registration effect.This paper proposes some solutions to the problems existing in multimodal image registration.The contents are as follows:(1)Aiming at the two problems of difficulty in extracting high repetition rate feature points on multimodal images and poor descriptor robustness,the paper proposes a multimodal image registration method based on max orientation gradient(MOG).This method starts with an anisotropic filter and constructs an anisotropic scale space(ASS)to replace the Do G scale pyramid in the SIFT algorithm,which can smooth the noise while retaining the edge structure information.In the feature description stage,a gradient definition method called MOG is proposed,which can better resist the distortion caused by local geometric and intensity changes between images,and can capture subtle structural changes.Based on MOG,a multineighborhood strategy is used to construct robust descriptors to achieve stable registration of multimodal images.(2)Feature-based methods are difficult to meet high-precision registration requirements.In order to achieve high-precision registration effects,region-based methods can be used.However,the region-based method cannot effectively deal with the geometric distortion between images.In addition,the local descriptors or similarity criteria in the region-based method are not robust and are susceptible to noise.In response to the above problems,this paper proposes a high-precision multimodal registration method based on a hybrid model.In the preregistration stage,the improved SIFT algorithm is used to roughly align the multimodal image pairs.In the fine registration stage,this paper uses anisotropic structure tensor to extract structural information and construct dense descriptors,which has good modal invariance and noise resistance.In template matching,using the parallelism of the tensor orientation as the similarity criterion can solve the intensity reversal phenomenon between multimodal images.In order to further strengthen the robustness of the similarity criterion,the tensor orientation parallelism and the gradient mutual information are combined to form a new similarity criterion to achieve high-precision multimodal image registration.