Research On SAR Image Speckle Removal Algorithm Based On Deep Learning

Synthetic Aperture Radar(SAR)image will be interfered by speckle noise,which makes it difficult for computer vision system to automatically interpret SAR data.Therefore,it is an essential step to remove speckle noise before applying SAR image to various tasks.Although a large number of SAR image denoising algorithms have been proposed,good results have been achieved.However,it is still a hot topic to remove speckle while maintaining more texture details of SAR image.Aiming at this problem,this paper starts from preserving the texture details of the image,and uses the powerful learning ability of deep learning neural network to study the SAR image speckle removal problem.The specific work of this paper includes:(1)A SAR image despeckle algorithm based on GAN network is proposed.Firstly,the generator network is used to pre despeckle the SAR image;The pre despeckle SAR image and clean SAR image are input into the discriminator;The discriminator feeds back to the generator by judging whether the image is true or false,prompting the generator to adjust the parameters accordingly,and encouraging the generator to generate a cleaner image;Until the generator and discriminator reach a Nash balance,the network training is completed.When GAN network is used to train SAR images,the parameters of the generator are updated not directly from the data samples,but by using the back propagation from the discriminator,which can obtain better generalization results.In the local network of the generator,residual dense blocks are used to extract SAR image features as much as possible,so that shallow features and deep features can be reused,and residual connection is added between residual dense blocks to maintain the edge information of the image to the greatest extent.Experiments show that this method can get better visual effect SAR image and higher objective numerical index.(2)A SAR image despeckle algorithm based on wavelet transform and neural network is proposed.Firstly,logarithmize the SAR image,and then use wavelet transform to decompose the SAR image at the first level;input the four wavelet subbands into the neural network for training,and finally pass the four subbands output through the network training through wavelet inversion.Transform and image indexation to get the SAR image after speckle removal.For image logarithm,the multiplicative noise model can be changed into the additive noise model to adapt to the residual connection.Using the wavelet coefficient matrix instead of the original image as the input of the neural network can not only remove the noise spots in different subbands,but also improve the sparsity of the input data,so as to compress the mapping range of the network and reduce the difficulty of training.In addition,the network uses the interval dense connection to reuse the shallow features of the image.The proposed method can effectively remove the speckle and retain a lot of texture information of SAR image.