Research On Despeckling Algorithm For SAR Image Based On Multi-scale Convolutional Neural Network

Synthetic aperture radar(SAR)is a coherent imaging system that can yield microwave remote sensing with high-resolution images without being limited by weather conditions and time.SAR images have the advantages of multiband,multipolarization,strong penetration,and all-weather acquisition.They are widely used in marine resource monitoring,military applications,forest monitoring,and other fields.However,due to coherent imaging mechanisms,serious multiplicative speckle noise exists in SAR images.The speckle artifact breaks down the edge of the target in SAR images,which further makes image understanding and analysis more difficult while creating considerable difficulties in target recognition and classification.Therefore,eliminating speckle noise in remote sensing plays a vital role in its application.With the continuous development of machine learning,methods based on deep learning have achieved good results in image processing(such as denoising,deblurring,etc.).The coherent noise suppression algorithm based on deep learning is mainly the application of Convolutional Neural Network(CNN)in coherent noise suppression.Based on the observation that the deep learning method can effectively model the image prior by learning a direct mapping function and the sparseness is a key characteristic of image,the paper takes the advantages of the discriminative learning and the sparseness of SAR images together,and proposes two algorithms for despeckling based on multi-scale convolutional neural networks.The main research work of the paper is as follows:(1)Speckle removal using hybrid frequency modulationsThe paper combines multi-scale geometric transformation and proposes a new despeckling algorithm.First,NSST is applied to a noisy SAR image to gain low-frequency and high-frequency coefficients.Second,a learned deep CNN model and CCS are applied to eliminate the speckle noise in the low-frequency coefficients and the high-frequency coefficients.And then the denoised low-frequency coefficients and high-frequency coefficients are obtained.Finally,the denoised image is obtained by applying inverse NSST to the denoised coefficients.Compared with existing algorithms,the results of the experiment indicate that the method has a better despecking effect,which can effectively remove speckle and maintains more detailed information retention.(2)SAR despeckling based on Shearlet transform CNNTo make better use of the sparse prior knowledge of images,the paper proposes a new despeckling algorithm combined with deep learning.First of all,NSST is applied to a SAR image.Then,a pre-trained ST-CNN denoiser is adopted to the images to obtain denoised sub-images.Finally,the final denoised image is obtained by applying inverse NSST to the denoised images.Experimental results show that the algorithm can effectively suppress coherent noise and retain the detailed information of the image.