SAR Image Super Resolution Based On Learning Strategy

High quality synthetic aperture radar(SAR)image has important application value in many fields,however,due to the high cost of hardware and the interference of the immobile factors in the imaging process,the SAR image obtained is often unable to meet the actual needs.The traditional method of promoting the SAR resolution by improving the system and processing in the imaging phase is also limited.In recent years,the technology of super-resolution reconstruction based on learning methods has developed rapidly,and has achieved very good results in the field of optical image.Therefore,this paper will combine the characteristics of SAR images,apply the learning based image super resolution reconstruction technology to the field of SAR image super-resolution reconstruction.Besides,we will combine the image classification technology and the image super resolution technology from the point of view of practical application.Based on the National special support program for high-level talents of China(The interpretation and object identification of SAR image).This paper study super-resolution for SAR image based on machine learning,deep learning and Generative Adversarial Network.The main contents are as follows:1)In this chapter,combined with the characteristics of SAR images,a method of super-resolution reconstruction of SAR images based on cartoon texture decomposition and joint optimization is proposed.This method jointly learn multiple regressor to minimize all reconstruction errors of training data,and then choose optimal regression function for different input low resolution image blocks,which overcomes the problem of SAR image having more complex structure.At the same time,we add cartoon texture decomposition strategy to our method,which decomposes all SAR images into cartoon images and texture images and process separately.In the final reconstruction stage,we will create a complete SAR image from the high-resolution cartoon images and texture images.Experiments show that our method has better performance in suppressing the speckle noise of SAR images and the recovery of spot target.2)In this chapter,a SAR image super-resolution method based on perceptual loss and GAN is proposed.The loss of perceptual loss refers to the loss of the advanced features extracted from the pre trained network,which has a superior performance in restoring the image texture edge information.The discriminate model in the GAN forces the reconstructed image to be better conformed to the statistical distribution of the real SAR image.The experimental analysis shows that the SAR image reconstructed by our method has a superior result when the magnification is larger.3)This work of this chapter is to research the effect of image super resolution for image classification.First,we train a VGG network that classifies remote sensing images.Because of the lack of data,we used the migration learning method in training the VGG network of remote sensing image classification.We use the parameters of pre trained VGG network to initialize the first few layers of the network,randomly initialize last several layers,and fine tune the network when training Then we use the method in the fourth chapter to magnify images in order to explore the accuracy of image classification of this method compared with the traditional method.Experimental results verify the validity of the method.