| Deep learning technology utilizes large amounts of data to drive deep neural networks to automatically extract features,which enables better extraction and representation of high-level features from input data,and thus has been applied successfully in radar image processing.However,existing radar image processing methods based on deep learning mainly adopt a supervised learning strategy,which requires a large amount of images with ground truths during the network training.In practice,obtaining these ground truths is extremely difficult,and in some cases impossible.When a deep neural network is trained with ground truths obtained from simulations or traditional methods,the network performance will be limited by the realism of the simulations or the performance of the traditional methods.When the supervised deep learning technique is applied to radar image processing,it is difficult,if not impossible,to obtain ground truths for the training.This dissertation studies the research of radar image processing based on unsupervised deep learning,where ground truths are not neessary for the training.The main accomplishments and contributions are summarized as follows:1.We propose an image registration method based on unsupervised image-to-image translation for the challenging registration of SAR and optical image.The proposed method simplifies the registration of SAR and optical images into the optical image registration.Specifically,for a SAR and an optical image,the SAR image is first translated into a pseudo-optical image in an unsupervised manner,and then an existing optical image registration method is employed to estimate the affine transformation parameters between the pseudo-optical image and the optical image,which indirectly realizes the SAR and optical image registration.Furthermore,to improve the quality of pseudo-optical images generated from SAR images,the improved Cycle GAN is developed by modifying the generators,the discriminators and the loss function of the original Cycle GAN.Experimental results show that the proposed method significantly improves the registration performance of optical and SAR image registration.2.We propose an ISAR image super-resolution method based unsupervised deep learning.Compared to ISAR image super-resolution methods based supervised deep learning,the proposed method trains the super-resolution network in an unsupervised manner,which does not require high-resolution images as ground truths and thus is suitable for real world applications.The well-trained network directly produces high-resolution ISAR images in real time and thus avoids time-consuming iterative optimization,achieving a 100x speed-up compared to the super-resolution method based compressive sensing.Additionally,a pseudo l0-norm has been proposed as the sparse constraint for the exact image reconstruction.Processing results of the real ISAR data demonstrate that the proposed method has a good super-resolution performance with a very acceptable generalization ability.3.An unsupervised cross-scene detection framework is proposed for moving target detection in video SAR,which effectively combines the advantages of traditional video SAR detection methods and CNN-based object detection methods.The proposed detection framework mainly consists of an image registration network and an object detection network based on background removal.The image registration network is designed for compensating the background motion between the frame to be detected and its nearby frames in real time.It can achieve a 20x speed-up compared to the classical SAR image registration algorithm and has a good generalization ability due to its unsupervised training manner.Furthermore,the object detection network based on background removal eliminates the background of the image to be detected at the feature level,which makes the video SAR data of different scenarios similar.Consequently,the object detection network trained in one scene can be applied successfully to processsing the data in another in another scene completely unknown during the training.Finally,the effectiveness of the proposed detection method is verified by two sets of real video SAR data in different scenarios.4.A two-step registration framework based on unsupervised deep learning is proposed for high-precision image registration of video SAR,which gets rid of the dependence on ground truths during the training.The designed framework is a cascade of two convolutional neural networks.The first network recovers the global transformation between the reference and unregistered images.Then,the second network takes the reference image and the registered image from the first network as inputs and then predicts a deformation field.After that,we put a limitation on the predicted deformation field to prevent moving target shadows from being aligned.Finally,the deformation field with limitation is used to compensate local deformations between the two images.It is verified by the real video SAR data that the proposed registration method can effectively compensate the local deformation between video SAR images,and thus improves the registration accuracy and facilitates subsequent video SAR shadow detection.In addition,the proposed registration approach is verified to have a good generalization ability by using video SAR data of unknown scene during the training.5.We propose an unsupervised video SAR image denosing method to solve the problem that SAR images uncontaminated by noise are unavailable for the training in the real world.The porposed method can suppress noise in video SAR images while protecting moving target shadows from blurring.Experimental results show that the proposed denosing method can achieve a good tradeoff between the noise reduction and the texture structure preservation.In addition,the proposed approach shows satisfactory processing efficiency and a good generalization ability. |
