Research On SAR Image Denoising Algorithm Based On Convolutional Neural Network

Synthetic Aperture Radar(SAR)is a coherent imaging system which can overcome the shortcomings that optical and infrared systems have and produce high-resolution images of terrain and targets.Therefore,it is widely used in various aspects.With the rapid development of secience and teconology,SAR is required to provide higher resolution and clearer SAR images.Because coherent noise seriously affects the quality of SAR images and is harmful to the interpretation of SAR images and various applications,so suppressing speckle has become an important part of SAR image processing.Moreover,the quality of coherent noise suppression will directly affect the accuracy of quantitative analysis and application of radar images,it also further affects image processing effects such as image segmentation and edge detection.How to effectively suppress speckle has become a difficult and hot spot in SAR image research.In recent years,with the research on various reverse problems in low-level vision,scholars have found that model-based optimization methods and discriminative learning methods have become important strategies for solving such problems,including image denoising problems.Convolutional neural network is one of the discriminative models commonly used in deep learning.Deep learning-based methods have been widely used in image processing such as image denoising and super-resolution and have achieved good results.However,the image denoising algorithm based on convolutional neural network is only very effective for image denoising at a certain noise level,and cannot achieve blind denoising.This paper proposes two blind denoising algorithms based on convolutional neural networks for SAR images,combining guided filtering based fusion algorithms and noise level estimation algorithms.The main research work of the paper is as follows:(1)SAR image denoising algorithm based on convolutional neural network and guided filteringIn order to avoid the mismatch of noise levels that would seriously affect the denoising of SAR images,this paper proposes a new SAR image denoising algorithm based onconvolutional neural network and guided filtering image fusion algorithm.First,we use trained CNN prior denoisers which is fast and effective with different noise levels to denoise the SAR image and obtain five denoised SAR images.Then,guided filtering based fusion algorithms is used to fuse the five denoised images to obtain the final denoised image.This algorithm combines the advantages of the model-based optimization method and the discriminative learning method.It not only can suppress speckle like the model-based optimization method,but also has the advantage of fast and flexible like discriminative learning methods.Experimental results show that the proposed algorithm has better denoising performance than the existing denoising methods,it can effectively remove noise and significantly improve the visual effect of the image.(2)Blind denoising algorithm based on FFDNet denoising model for SAR imagesIn order to improve the shortcomings of the existing denoising algorithms that cannot adapt to all noise levels,this paper combine scale-invariant kurtosis and piecewise stationary noise level estimation algorithm with fast and flexible convolutional neural network,and proposes a new blind denoising algorithm for SAR images.Firstly,we estimate noise level for SAR image through the noise estimation algorithm based on scale-invariance kurtosis and piecewise stationary.Then we use the estimated noise variance as input parameters of the convolutional neural network to perform SAR image denoising,and finally achieve blind denoising of the SAR image.Experimental results show that the denoising effect of this algorithm is more obvious compared with the current popular denoising algorithms,and it can better maintain the edge and texture information of images.