Research On Speckle Suppression In SAR Images Based On Convolutional Neural Network

Synthetic aperture radar(SAR)is an active imaging radar with the advantages of working in all-time,all-weather and wide frequency band.However,the inherent defects of SAR imaging technology will inevitably generate coherent speckle noise,which will bring difficulties to the subsequent interpretation of radar images.Therefore,the SAR image despeckling has been an important topic for researching in the interpretation of radar images.Recently,deep learning technology has been gradually applied in various tasks of natural image processing and achieved excellent results,which provides new ideas for SAR image despeckling.This thesis researches on the SAR image despeckling algorithm based on convolutional neural network(CNN).The main innovations of this thesis are as follows:(1)A SAR image despeckling algorithm based on multi-scale interactive network is designed in this thesis.In this algorithm,the multi-scale interactive feature extraction module(MIFEM)is constructed by kernels with different sizes and residual connections,which can obtain features of different receptive fields and speed up the convergence of network;The network can make full use of shallow texture features by using simplified dense connection between MIFEMs.The experimental results show that the proposed algorithm not only uses less calculation parameters,but also ensures the improvement of performance.(2)From the view of improved activation function,a SAR image despeckling algorithm based on learnable activation function is proposed.As an activation function,the rectified linear unit(Re LU)function can make the network sparse and speed up the convergence of network,which is applied to various network models.The Re LU function from the perspective of threshold unit is reinterpreted,and the learnable activation function is designed based on the Re LU function.In order to verify the performance of the function,the existing algorithms with simple network structure such as ID-CNN are selected to avoid the influence of complex network structure on its performance comparison.The activation function is introduced into the MIFEM and named as x MIFEM.The experimental results show that this function improves the despeckling performance of the network at the expense of fewer parameters.(3)A SAR image despeckling network algorithm based on multi-level wavelet transform is proposed from the view of combining traditional algorithm and neural network.Wavelet transform with excellent time-frequency characteristics can switch images to frequency domain to achieve despeckling.In addition,scale decomposition and reversibility of wavelet transform is helpful to avoid losing lots of information in the process of down-sampling and up-sampling.Based on these advantages,the convolutional neural network is inserted into the multi-level wavelet transform and inverse wavelet transform.The concatenate operation is used to make full use of the shallow feature information of the network.Finally,the residual connection is introduced to obtain the network structure.The experimental results show that the proposed method can effectively improve the performance of despeckling.