Study On High-Resolution ISAR Imaging Methods Based On Deep Network

Inverse Synthetic Aperture Radar(ISAR)has the features of all-day,all-weather,long-range and high-resolution,and plays an important role in space situation awareness and airborne target surveillance.For non-cooperative targets,ISAR obtains high range resolution through large Time-Bandwidth Product(TBP)signals such as Linear Frequency Modulated(LFM)signals and pulse compression,and obtains high azimuth resolution by taking advantage of the virtual aperture formed by the relative motion between the target and ISAR.Under ideal observation conditions such as high signal-to-noise ratio and complete echoes,well-focused imaging results can be obtained through classical imaging algorithms such as the rangeDoppler algorithm or polar formatting algorithm after translational compensation.However,in practical imaging tasks,ISAR often faces complex observation conditions,such as low signal-to-noise ratio echoes,sparse observation scenarios,etc.,which lead to the rapid performance degradation or even failure of existing imaging methods.To address this problem,high-resolution ISAR imaging methods based on sparse reconstruction have made great progress in recent years.However,the existing sparse reconstruction imaging methods have the following problems:1)the optimal parameters of the algorithms usually vary with the defect form and signal-tonoise ratio,and require manual search rather than automatic selection;2)the related algorithms usually have a large number of iterative steps,with slow convergence speed and longtime consumption.Since the high-resolution imaging method combined with deep networks has the advantages of high reconstruction accuracy,fast computational speed and no parameter tuning,it provides a new idea to solve the high-resolution ISAR imaging problem under complex observation conditions.With the support of the National Natural Science Foundation of China(NSFC)key programs and general programs,the dissertation is closely related to the difficult and critical problems in practical imaging tasks,and carries out in-depth research on high-resolution ISAR imaging methods based on deep networks.The main contents include: 1)To address the problems of traditional sparse signal reconstruction methods that require manual tuning of parameters,etc.,a model-driven deep network imaging method is studied to achieve high-resolution focused imaging under sparse observation conditions;2)To address the problem that model-driven deep networks are sensitive to noise,a data-driven deep network imaging method is studied,which achieves noise-robust sparse high-resolution imaging;3)To address the problem that existing imaging methods rely on training data and have poor generalization performance when the test data differs significantly from the training data,a high-resolution imaging method based on deep reinforcement learning is investigated,which significantly improves its generalization performance.4)To address the problem that existing methods are sensitive to different loss rates,a deep network imaging method combined with a hypernetwork is investigated,which enables robust imaging under different loss rate conditions.The specific research of this dissertation can be summarized in the following four parts.1.Sparse observation high-resolution ISAR imaging combined with deep unrolling networks Although the sparse signal reconstruction algorithm provides an effective way to solve focused imaging under complex observation conditions,its imaging performance is very sensitive to internal parameters,and the optimal parameters need to be manually adjusted,which is time-consuming.To address this problem,this dissertation first models the sparse observation of the echo signal and gives the iterative formulation of the sparse signal reconstruction algorithm two-dimensional alternating direction method of multipliers(2D ADMM);Then,the 2D ADMM algorithm iterative steps are expanded into a finite layer deep network by using the deep unrolling technique to address the problems such as the difficulty of internal parameter tuning,and the internal adjustable parameters are set as network parameters;finally,the optimal parameters are automatically learned through network training,and the high-resolution ISAR imaging method based on 2D ADMM-Net is proposed.The processing results of simulated and measured data show that the trained network can efficiently achieve high-resolution ISAR imaging under sparse observation conditions.2.Data-driven network-based signal-to-noise ratio robust high-resolution ISAR imaging In response to the problem that deep unrolling network imaging methods are sensitive to noise and require separate training of the network for different signal-to-noise ratio echoes,resulting in high temporal and spatial complexity,we conduct research on data-driven network-based signal-to-noise ratio robust high-resolution ISAR imaging.Firstly,highresolution ISAR imaging method based on UNet++ is proposed to achieve two-dimensional high-resolution imaging robust to different signal-to-noise ratios under complete echo conditions.Further,a high-resolution ISAR imaging method based on plug and play twodimensional alternating direction method of multipliers network(PAN)is proposed for incomplete echoes in combination with plug-and-play technology.The proposed method PAN combines the sparse signal reconstruction algorithm 2D ADMM and the deep denoising convolutional neural network(Dn CNN)to achieve robust and accurate reconstruction of target scattering points under different signal-to-noise ratio conditions.Finally,the effectiveness of the proposed method is verified by simulated and measured data.3.High-resolution ISAR imaging based on deep reinforcement learning Due to the constraints of observation conditions,it is difficult to obtain a large amount of measured data for some targets,so existing deep imaging networks usually use random point simulated data for model training,which will lead to the degradation of the imaging performance of measured data when the distribution of measured data is different from that of simulated data,i.e.,poor generalization performance.To address this problem,a highresolution ISAR imaging method based on Swin RL is proposed.The method models the sparse signal reconstruction algorithm 2D ADMM as a sequential decision problem,models the internal adjustable parameters as actions,and uses Swin Transformer as the backbone network of the policy network and value network,and then guides the optimization of the actions,i.e.,the internal adjustable parameters of the 2D ADMM algorithm,through network training in a reinforcement learning framework.The imaging results of simulated and measured data show that compared to existing deep unrolling networks with fixed parameters,trained agent can automatically give optimal internal parameters according to different input data,and then well-focused imaging results can be obtained by executing 2D ADMM algorithms with optimal parameters.4.Loss rate robust high-resolution ISAR imaging combined with hypernetwork To address the problems that existing deep network imaging methods lack robustness to defect forms and require retraining the network when the defect form changes,resulting in high time and space complexity,a 2D ADMM algorithm independent of the loss rate is first derived by keeping the dictionary dimension constant,and then a high-resolution ISAR imaging method based on CSAN is designed.The method uses CSWin Transformer for feature extraction of the defocused RD image,and outputs the optimal parameters corresponding to the loss rate through network training,and then obtains well-focused imaging results by executing the 2D ADMM algorithm with the optimal parameters.The imaging results of the simulated and the measured data show that the model can achieve robust focused imaging of the target under different loss rate conditions.