Optical And SAR Images Based On Feature Decoupling Network Reasearch On Registration Methods

Image registration of optical and synthetic aperture radar(SAR)images is an important task that can fuse disparate information obtained from these two different sensors to achieve more comprehensive and accurate geo-information.However,due to the problems such as speckle noise,nonlinear radiometric differences,and computational complexity,the accuracy and efficiency of image registration of optical and SAR images are limited.Feature decoupling is the process of decomposing feature representations into multiple independent factors by removing redundant information within multiple correlated features to obtain more independent feature representations.Feature decoupling is commonly used for tasks such as image classification,object detection,and image reconstruction.This paper applies the idea of feature decoupling to the feature extraction process of optical and SAR image registration and proposes a novel method with the following main research contents and innovative ideas:(1)A registration method based on the feature decoupling network(FDNet)is proposed.FDNet consists of two parts: residual denoising network(RDNet)and pseudo-siamese fully convolutional network(PSFCN).This method overcomes the disadvantages of both feature-based and region-based methods in terms of registration accuracy and computational complexity by using a fast template matching algorithm.Specifically,this algorithm uses FAST feature point detection to generate the center of the initial template,uses PSFCN to extract local feature descriptors and match the initial template,and then searches for the optimal matching template in a small search window around the matched initial template.(2)The feature decoupling strategy based on RDNet is studied.This paper designs a statistical model for coherent speckle noise in SAR images using RDNet and defines a loss function based on mean squared error(MSE)and total variation(TV)to propagate coherent speckle noise.Finally,PSFCN and RDNet are used to learn deep representations of semantic and noise information,respectively,and decouple them at the interval convolutional layer.(3)A template size adaptive selection strategy based on twodimensional entropy is proposed,which can select an appropriate template size according to the content richness of SAR images.On a publicly available registration dataset,this paper demonstrates the robustness of the proposed method to SAR noise and achieves better registration accuracy than other state-of-the-art methods.