Research On Deep Learning-based Sparse SAR Imaging And Despeckling Integrated Processing

Synthetic aperture radar（SAR）has been widely applied in military and civilian fields owing to its ability to offer the high-resolution imaging results of the monitoring area in all-time and all-weather conditions.However,the matched filter-based SAR imaging method requires a huge amount of measurement data according to the Nyquist rate when offering the high-resolution imaging results.This demands higher quality of the hardware for data measurement,transmission,and storage in SAR system.Besides,modern radar requires SAR system possessing multiple imaging models.Thus,the imaging interval and resource for single monitoring area is limited.This would lead to the discontinuity or loss of the sampled data for this area,degrading its imaging quality.On the other hand,due to the coherent imaging mechanism in the SAR system,SAR images inevitably suffer from the speckle noise.This speckle noise does not only reduce the quality of the SAR image,but also interferes with the subsequent detection,recognition,interpretation and segmentation.To address these two problems,researchers proposed an optimization model-based reconstruction method according to the statistical properties of the complex-valued SAR images,trying to recover the speckle-free backscattering coefficients from the downsampled measurement data directly and realizing SAR imaging and despeckling.However,this method is established on the universal SAR imaging observation matrix model.Within solving procedure,this method recovers SAR image in vector form,which induces excessively high computational cost.Thus,this method may not be practicable for real-time SAR imaging and despeckling processing.In recent year,deep learning technology has been rapidly developed and it has achieved great breakthroughs in downsampled data imaging and image denoising for natural image processing,which bringing novel ideas for sparse SAR imaging and SAR image despeckling via downsampled data.To address the above problems,this thesis,combining the deep learning and radar signal processing technology,investigates the deep learning-based sparse SAR imaging and speckle removal for downsampled data.Its main contributions and innovative achievements are as follows:（1）Aiming at ignoring the despeckling task in the sparse SAR imaging of the traditional radar signal processing,this thesis proposes an observation model considering the sparse SAR imaging and SAR image speckle reduction at mean time,called the sparse SAR imaging-despeckling observation model,and proposes a deep network for simultaneously realizing the sparse SAR imaging-despeckling.Assuming the noise in the proposed observation model obey the complex Gaussian distribution,an optimization model with a L₂norm as the fidelity term is established by utilizing the maximum a posteriori（MAP）estimation.In addition,two L₁ norms are employed as the constraints for the sparse SAR imaging and SAR image speckle reduction tasks to improve the imaging and despeckling quality,respectively.The former is for the sparse constraint.The latter consisting of a convolutional neural networks（CNN）-based projection operator is for the constraint of image detail preservation.The complex-valued split Bregman method（CV-SBM）is adopted to solve the problem by separating the original one into several easily-solving subproblems.Then the iteration steps for solutions are transformed into a parameter-learnable and architecture-fixed deep network.The experimental results with the simulated and real data demonstrate the validity in sparse SAR imaging and speckle suppression of the proposed deep network.（2）Due to the low quality and high computational complexity of the deep network for the sparse SAR imaging and despeckling,a real-time deep network is devised.To fit more diverse imaging regions,we suppose that the noise in the SAR imaging-despeckling observation model follows complex generalized Gaussian distribution.Based on this assumption,an L_q-norm（q>1）as the fidelity term is exploited to improve the optimization model in 1.Within the solving process by using the CV-SBM and transforming it into a deep network,a computationally efficient solution is presented for the fidelity term-related subproblem,due to the specific down-sampled strategy in SAR system,and a substitutive cost function and a CNN structure are introduced to solve the projection-related subproblem,improving the despeckling ability.Numerical experiments based on simulated and real data validate that the proposed deep network could promote the image reconstruction quality and the real-time capability of the sparse SAR imaging and despeckling for downsampled data.（3）To further improve the image reconstruction quality of the sparse SAR imaging and SAR image speckle reduction,a deep work based on sparse,low-rank,and deep CNN priors is devised.The SAR image in the SAR imaging-despeckling observation model is split into a sparse matrix and a low-rank matrix.Based on this information,an optimization model with an L₁ norm and a weighted nuclear norm as the sparse and low-rank constraints is established.Moreover,a deep CNN prior regularization is introduced to further promote the speckle removal ability.Finally,a deep network is proposed to solve this optimization problem.Experimental results demonstrate the proposed deep network could promote the image reconstruction quality for the sparse SAR imaging and SAR image despeckling.