Research On SAR Image Processing Based On Deep Learning

Since SAR has advantages of all-time and all-weather,it has great application value in military and civilian fields.However,SAR images have inherent speckle noise,which seriously affects the effective implementation of subsequent operations such as image segmentation and target detection.It is of great significance to study on SAR image denoising and target detection techniques.Recently,deep learning technology has developed rapidly and various excellent models have emerged in an endless stream,which makes excellent achievements in the field of natural image processing,providing new ideas for SAR image denoising and target detection.This thesis researches on the SAR image denoising and target detection based on deep learning.The specific work and innovations of this thesis is as follows:The basic structure and related algorithms of convolutional neural networks are introduced in detail.And the characteristics of different activation functions and cost functions are analyzed,the forward propagation and back propagation algorithms are introduced.In addition,this thesis describes four wildely used convolutional neural networks,including the residual module of ResNet and VGG-16 network model,providing a support for SAR image denoising and target detection research.This thesis studies on SAR image denoising based on deep learning.Firstly,the imaging mechanism and model of SAR image speckle noise are analyzed.Two classical denoising methods and one deep learning-based network are briefly introduced.In order to improve denoising performance and reduce computational complexity,this thesis makes some improvements on SAR-DRN.This thesis builds an improved denoising network with fewer layers and parameters,which can effectively reduce the amount of calculation and the memory usage of the computer.Then,this thesis simulates SAR images with different intensity speckle noise characteristics using natural images to train the network.Finally,the denoising performance of our network is tested using simulated and real data.Compared with three existing methods,our denoising network has advantages in denoising effect and calculation time.This thesis studies on target detection of SAR image based Faster R-CNN and explores an improvement measure.Faster R-CNN is a state-of-the-art algorithm based on deep learning,which has advantages of high speed and precision,and is excellent in the field of natural image processing.Firstly,the structure and training process of Faster RCNN algorithm are analyzed.Then,MSTAR dataset which has sufficient amount of data is applied to verify the effectiveness of the algorithm on SAR image target detection.However,MSTAR data has a single target and a simple background.To further verify the performance of the Faster R-CNN algorithm for SAR images with complex background and diverse targets,this thesis combines target extracted from MSTAR image and with three different real SAR images and produces the simulation dataset with different backgrounds and multiple targets.This thesis tests the performance of Faster R-CNN algorithm on the simulated data and analyzes the effects of complex background and multi-target on the detection results.Experimental results show that Faster R-CNN algorithm has a good detection performance for most of the targets of the simulated data,but some targets has bad detection precision.This thesis analyzes the possible reasons and concludes that complex background makes it more difficult to distinguish those objects with similar structures.Aiming at this problem,this thesis explores an improvement measure.This thesis extracts targets with low detection accuracy in terms of location information,and then designs a network for secondary classification,thus improving the detection precision.