Research On SAR Image Denoising And Segmentation Methods Based On Deep Learning

As a remote earth observation technology,microwave remote sensing technology plays an important role in many fields closely related to national development,such as meteorological observation,ocean detection,disaster prevention and so on.As a high resolution remote imaging sensor commonly used in microwave remote sensing,Synthetic Aperture Radar(SAR)has unique advantages in remote sensing observation of the earth,and has been widely used since its advent.However,due to the inherent nature of SAR imaging system and the complexity of the observation target,SAR images are less visualized than optical images,and are vulnerable to speckle noise,perspective contraction,shadow and so on.How to effectively extract and process information from SAR images has become an urgent problem.In recent decades,although some SAR image processing algorithms have been proposed,they mostly rely on the characteristics of manual design,and there are some problems in accuracy and robustness.With the development of in-depth learning technology,convolutional neural network algorithm has made great breakthroughs in various fields of computer vision,which also provides a new idea for SAR image processing research.In this paper,a SAR image denoising and segmentation algorithm based on depth learning is proposed.The main contents and achievements of this paper are as follows:1.A SAR image denoising method based on depth residual learning in transform domain is proposed.The principle and mathematical model of SAR speckle noise imaging are discussed.Then a SAR image denoising network is proposed,which follows the idea of residual learning and learns the difference between the noise level of input SAR image and that of real SAR image in transform domain.Finally,the model is applied to the denoising of synthetic image and real SAR image.The results of subjective vision and objective indicators are given.2.A SAR image segmentation method based on multi-scale feature fusion is proposed.The main ways of fusing multi-scale online text features in convolutional neural network are analyzed.On this basis,the network results of the model are given.The network is divided into three main parts: encoder,decoder and multi-scale feature fusion module.An experiment on road target segmentation in SAR image is carried out to verify the effectiveness of the proposed algorithm.The experimental results show that the SAR image denoising and analysis algorithm based on depth learning proposed in this paper can effectively reduce the speckle noise of SAR image,improve the subjective effect and objective quality of the image,and achieve good results in road target segmentation of SAR image.The algorithm is based on a large number of data training.It can capture the deep structure information and semantic features of the image.It has good robustness and accuracy,and has certain research and engineering application value.