SAR Image Enhancement Based On Deep Neural Network

Due to its ability of all-day and all-weather observation for the region of interest,synthetic aperture radar（SAR）has been widely applied in military and civilian areas such as battlefield reconnaissance,marine exploration,agriculture investigation,etc.As a key technology of SAR image processing,deep neural networks can model linear and non-linear systems,and learn target features automatically from images according to require-ments,with high accuracy,flexibility,and versatility.SAR image processing is the basis of SAR data understanding and interpretation,which usually includes image enhancement,fusion,segmentation,target recognition,tar-get detection,and tracking.This dissertation focuses on the study of SAR image enhance-ment algorithms using deep neural networks and the main innovations are as follows:（1）The spatial characteristics of deep neural networks are studiedAccording to the spatial retention characteristic of the deep neural network,the“match-ing criterion”which means that the input of each layer of the neural network and the net-work parameters of that layer satisfy a one-to-one correspondence in space is proposed.Based on this criterion,a kernel mapping network（KM-Net）that can accomplish both target classification and angle estimation is designed.Different from the general network that needs to be trained,the weight parameters of KM-Net are obtained by performing the kernel mapping rotation transformation on the weight parameters of the reference net-work.In order to improve the classification and angle estimation ability of KM-Net,the shifting-pixel rotation method and the octagonal convolution kernel are designed.The shifting-pixel rotation method can effectively solve the mismatch problem caused by in-terpolation rotation by moving the weight parameters of each position in the convolution kernel.The octagonal convolution kernel has the characteristic of 45°rotation symmetry,which can solve the mismatch problem of 45°rotation of the standard rectangular con-volution kernel.A kernel rotation augmentation network（KRA-Net）is designed to solve the problem caused by insufficient training samples.The network realizes multi-angle augmentation for the internal features of the network by rotating the convolution kernel during training and can improve the classification accuracy when the training set is small.While solving the problems of missing SAR target angle and reduced recognition rate when there are fewer training samples,the research provides theoretical guidance for the design of SAR image enhancement and shadow tracking networks.（2）Despecking and auto-focusing networks are designedA multi-scale recurrent network（MSR-Net）for SAR image despecking is proposed.The network adopts the strategy of“coarse to fine”,and invokes the same network mod-ule in a cascaded way,which directly suppresses speckle noise and generates noiseless output.Experimental results show that the MSR-Net has excellent denoising ability and can achieve the state-of-the-art results on both simulated and real SAR images with low computational costs,especially when the noise is strong.A convolutional long short-term memory（Conv LSTM）module is added to MSR-Net which imposes the network with the transmission and fusion abilities of multi-scale features.Meanwhile,a sub-pixel convo-lution unit is adopted to restore the resolution of the feature map to improve the efficiency of network.Two denoising algorithm evaluation metrics edge feature keep ratio（EFKR）and feature point keep ratio（FPKR）are proposed.They both can quantitatively calculate the degree which the denoising algorithm retains the typical features such as edges,an-gles,and textures.Aiming at the problem of image defocusing caused by the unknown position error of moving platform,a SAR image auto-focusing network（AF-Net）based on the encoder-decoder structure is proposed.This network can learn the geometric dif-ference in the image before and after the focusing processing,and achieve auto-focusing in the amplitude domain.Compared with the traditional phase gradient autofocus（PGA）algorithms,this network has the advantages of easy implementation and high efficiency,and can provide a new idea for the research of related problems（3）The shadow tracking and moving target refocusing methods are studiedA refocusing framework for the ground moving target is proposed according to the characteristic that the target’s shadow can reflect the true position of the target,which effectively combines video-SAR,shadow tracking network,trajectory optimization net-work,and moving target back-projection imaging algorithm.The simulation data are used to verify the effectiveness of the above methods and high-resolution images of ground moving targets with complex trajectories are obtained.A tracking network m Re³based on region searching strategy is designed.The m Re³can achieve high tracking accuracy of moving target shadow in video-SAR by effectively combining the convolutional and recurrent neural network,and extracting the target features while mining the temporal re-lationship of the target in adjacent frames.To improve the smoothness of the tracking trajectory,a long short-term memory based TV^β-LSTM is designed.Also,a high-order total variance loss L_TVβwhich can improve the network’s performance is designed.The center distance error of the trajectory can be less than one pixel after the optimization of the TV^β-LSTM.These studies can improve the reconnaissance and observation ability of SAR system and enhance the depth and breadth of the application of deep neural network in SAR image processing,which have strong theoretical significance and engineering value.