SAR Image Super-Resolution Reconstruction Based On Joint Discrimination Of High-Resolution And Low-Resolution Images

High-resolution SAR(Synthetic Aperture Radar)images have important applications in image classification and target recognition.At present,the generative adversarial network is widely used to improve the resolution of images.Through the discrimination of high-resolution images,the generated high-resolution images can have more high-frequency texture information.However,the existing image super-resolution reconstruction algorithms based on generative adversarial networks only include the discrimination of high-resolution images while ignore the role of low-resolution images.This does not guarantee that the generated high-resolution images can be accurately downsampled to the original low-resolution image.In order to make full use of the role of low-resolution images,a SAR image super-resolution reconstruction algorithm based on the joint discrimination of high-and low-resolution images is proposed.The joint discrimination algorithm of high-and low-resolution images adds the discrimination of low-resolution images based on the discrimination of high-resolution images.The discrimination of high-resolution images makes the generated high-resolution images contain more high-frequency texture information,and the discrimination of low-resolution images ensures that the reconstructed high-resolution images can be accurately downsampled to the original low-resolution images.Aiming at this problem,the main research work of this thesis is as follows:(1)A SAR image super-resolution reconstruction algorithm based on teacher-student discrimination is proposed.Aiming at the problem of inaccurate discrimination of the low-resolution images,the network architecture of teacher-student discriminator is used in this algorithm.First,the teacher discriminator is used to discriminate high-resolution images to ensure that the generated high-resolution images have more high-frequency texture information.Secondly,the student discriminator is used to discriminate low-resolution images to ensure that the reconstructed high-resolution images can be accurately downsampled to the original low-resolution images.Finally,the teacher discriminator is used to guide the training of the student discriminator,which improves student discriminator.(2)A SAR image super-resolution reconstruction algorithm based on multi-scale discrimination is proposed.Aiming at the problem of inaccurate single-scale discrimination space,a multi-scale discrimination method is used in this discriminator.The output of the discriminator is the joint probability that the image is real in different scale feature spaces,which further enhances the accuracy of discrimination.The multi-scale high-resolution discriminator is used to discriminate high-resolution images,and the multi-scale low-resolution discriminator is used to discriminate low-resolution images.In addition,in order to make the high-resolution discriminator and the low-resolution discriminator play different roles in the training process,an adversarial loss function based on the mean square error is proposed to make the role of the discriminative loss of the two images with different resolutions constantly changes with the training of the network.The proposed algorithms are tested on the Sentinel-1 dataset and compared with the traditional methods and deep learning methods.From the perspective of objective evaluation indicators and subjective effects,the proposed teacher-student discrimination algorithm and multi-scale discrimination algorithm perform better in reconstructing the texture details of high-resolution SAR images.