Research On SAR Image Quality Enhancement Method Based On Deep Learning

Since SAR has advantages of all-time and all-weather,it has research significance and great application value in military and civil fields.However,in the process of SAR imaging,the inherent coherent superposition of echoes,the lack of imaging resources,the radar system error and the relative movement of the SAR platform lead to the quality degradation of the SAR image,which is mainly represented by the speckled noise and the reduction of the resolution,which seriously affects the information extraction and interpretation of SAR image.Deep learning has developed rapidly in recent years and made many breakthroughs in the field of image processing,which provides new ideas for further quality enhancement of the imaged SAR images.In this thesis,combined with the currently public datasets and simulation datasets,the deep learning method is used to study the SAR image quality enhancement technology.The special work and innovation points are as follows:1.The degradation mechanism of SAR image is deeply studied,and the degradation model of SAR image quality is deduced.The solutions of numerical method and deep learning method are introduced and compared,which provide theoretical support for SAR image quality enhancement research based on deep learning.2.An improved deep convolutional network for SAR image denoising(PID-DNN)is proposed.Based on ID-DNN,reversible down-sampling and convolution kernel size reduction are adopted to greatly reduce the number of parameters.By introducing noise level map estimation network,the prior information is increased while the generalization capability of the network is improved,the performance and computational efficiency are improved.3.The SAR image residuals deep convolutional neural network(SRDRN)is proposed.Based on SRCNN modular design idea,combined with residual block technology and skip connection technology,the forward transmission of low-dimensional features,the extraction of high-dimensional features and the morphological correction in the process of super-resolution reconstruction are realized,which solves the problems of training difficulty and morphological distortion in SRCNN,and significantly improves the performance of super-resolution.The theories and methods mentioned in this thesis have been verified by simulation and measured data,and the results show that these proposed methods can effectively enhance spaceborne SAR image quality.