SAR Image Registration Based On Deep Learning

With the country’s strong support for aerospace technology,synthetic aperture radar technology has been continuously developed,and people can obtain a large amount of SAR image data from various platforms to meet the needs of military and civilian fields.The purpose of SAR image registration is to convert two images of the same scene captured from different times,different viewing angles or different sensors to the same coordinate system.The accuracy of SAR image registration directly affects the performance of subsequent image processing tasks.In recent years,deep learning has been introduced into SAR image registration to obtain some good progresses.However,due to some special properties of synthetic aperture radar images,there are some problems for the deep learning-based SAR image registration,such as difficultly constructing a mass of labeled training samples,the use of a single size makes the features extracted by the network not rich,speckle noise in SAR images resulting in negative impact on registration performance,etc.Based on these problems,this thesis further explores and studies for the deep leaning-based SAR image registration.The completed works mainly include the following points:1.For the SAR image registration task,most DL-based methods usually utilize matched and unmatched image patches to build training models.Generally,a fixed scale is set to intercept the image patch to generate the training set.However,the image patches of different scales contain different information,which can affect the performance of the registration.Moreover,it is difficult for the SAR image registration task to obtain a large number of labeled training samples.Based on this,this thesis proposes a multi-scale fusion SAR image registration framework based on deep forest.The method first constructs a training sample set of multiple scales,then using the deep forest as the basic learning model to train multiple matching models,and designing a multi-scale fusion strategy to integrate multiple prediction results,and registering the images.Finally,experiments are carried out on multiple datasets,and the registration results are visualized and analyzed to verify the effectiveness of the algorithm.2.Considering that it is difficult to obtain a large number of labeled datasets for SAR image tasks,this thesis proposes an adaptive SAR image registration method based on contrastive learning.In most SAR image registration methods,the training set and the testing set are usually constructed based on different images.However,there are obvious noise and spatial differences between two SAR images,so that the constructed training and test samples also have large differences,which results in a negative impact on the learning model.Therefore,the proposed method transforms the key points on the sensed image to the reference image through the registration transformation matrix,where only the reference image is used to construct training and testing sample.Then,the contrastive learning model is utilized to obtain better feature representations which are applied to calculate the image registration matrix model.It is worth noting that the constructed sample set is adaptively updated during the model training process,which can make the constructed matched sample pairs more similar.Moreover,a new evaluation index is proposed to confirm the accurate registration.The experimental results validate that the method has high registration accuracy and certain anti-noise.3.In the task of SAR image registration,most existing methods consider the image registration as a two-classification problem to construct the pair training samples for training deep learning models,whereas it is difficult to obtain a mass of pair SAR images.Based on this,from the view of multi-classification,the thesis takes each key point obtained on the SAR image as a category in the classification task,and it proposes a two-network transformed SAR image registration method based on transformer.The proposed method obtains key points from the reference image and the sensed image to construct the training set and the testing set,respectively,and then it designs the coarse registration strategy and the refined registration strategy to obtain the final matched points which are predicted as the same class.Finally,the transformation matrix is calculated based on the final matched points.Specially,considering that the key points from the reference image and the sensed image are inconsistent,this paper designs a two-network transformation model for SAR image registration,and it obtains the matched pair points which are consistent in two feature spaces trained by two networks,respectively.Experimental results and analyses illustrate the proposed method performs better registration performance in the task of SAR image registration.