Research On SAR Image Super Resolution Based On Generative Adversarial Network

High-resolution SAR(Synthetic Aperture Radar)images have important practical values in research and application fields,such as satellite remote sensing,disaster monitoring and so on.However,the low imaging resolution of SAR images makes it difficult to meet application requirements in practice.Traditional methods for super-resolution reconstruction of SAR images have high requirements for method model and prior knowledge,otherwise the performance of reconstruction methods is not satisfied.In order to release these requirements,the unsupervised generative adversarial network model was proposed.It can provide good reconstruction results by training and learning the relationship between high-and low-resolution SAR images,without prior knowledge.However,the game-based adversarial methods often cause "artifacts" in resulting reconstructed SAR images,leading to inconsistencies in image structure and content compared to reference images;moreover,the details,such as edge and texture,in reconstructed SAR images are insufficient to meet the practical application requirements.In order to overcome these problems,the research works based on the generative adversarial network model for improving the resolution of reconstructed SAR images are proposed in this thesis.The details are as follows:(1)A structure-enhanced SAR image reconstruction algorithm based on double sampling mechanism is proposed in this study,to deal with both "artifact" and inconsistency of structure and content problems in reconstruction of SAR images,which are created by the generative adversarial network model.This method firstly extracts low-level features from the input SAR image with a small-scale convolutional layer,then obtains the input features using a RRDB module,and finally reconstructs the high-resolution SAR image by alternately implementing the nearest neighbor interpolation(NNI)and sub-pixel convolution(SPC)methods.On the hand,the structural loss function is introduced to combine with the structural similarity(SSIM)evaluation index,for maintaining the structure and content consistency in the reconstructed and reference images.Results show that this alternate double sampling mechanism is benefit for the interactive fusion of feature information that can improve the reconstruction performance,and the new proposed structural loss function can alleviate the "artifact" in the resulting reconstructed image.(2)A method for providing more image details,such as edge and texture,as well as maintaining the structure and content consistency in reconstructed SAR images is proposed in this study.This method introduces a multi-scale receptive field module based on the original model,which improves the performance on the fusion of feature information with different sizes.Furthermore,in order to reduce the computational cost for speeding up the reconstruction process,this method uses a variety of cascaded small-scale convolution kernels to replace the large-scale convolution kernel.Results show that this method can both reduce network parameters and enhance the ability for detail feature extraction.To evaluate the effectiveness of the proposed methods,the algorithm models are trained with the Sentinel-1 data set and tested with high-resolution Terra-SAR images.Experimental results show that compared with traditional interpolation algorithms and classic deep learning methods,the methods proposed in this thesis have a significant improvement in qualitative visual performance and quantitative evaluation indicators that the structure and content in reconstructed images are consistent with reference images,and the information of details is more prominent.