SAR Image Speckle Suppression Method Based On Convolutional Neural Network

Synthetic Aperture Radar(SAR)can penetrate clouds,haze,dust and other climatic conditions,and has all-day and all-weather imaging capabilities.Therefore,SAR data has become an important source of remote sensing data.However,as a coherent imaging system,SAR inevitably produces granular speckle noise in its image,which greatly affects the image quality and hinders its application in many fields such as land use,agriculture,topographic survey,meteorological prediction and environmental monitoring.Therefore,it is very important to suppress speckle in SAR images.Traditional SAR image speckle suppression methods have achieved good results in speckle suppression performance.However,the running time and speckle suppression ability of these methods mainly depend on the selection of algorithm parameters,and there is a strong correlation between the parameters.It is difficult to adjust and the suppression effect is insufficient.The convolutional neural network has nonlinear learning ability and shows good potential for SAR image speckle suppression.Through the constructed network model,a large amount of data is trained and learned,and the nonlinear end-to-end mapping between SAR image and its speckle-free noise reference is realized to achieve the purpose of speckle suppression.Compared with traditional algorithms,this method can better capture the characteristics of speckle noise by establishing a nonlinear model and achieve more accurate speckle suppression effect.Large-scale SAR image data can also be used for training,thereby improving the robustness and generalization ability of the algorithm.Therefore,this paper proposes a SAR image speckle suppression method based on convolutional neural network.The main conclusions are as follows:(1)In this paper,the SAR image data set is constructed.In the SAR speckle suppression method based on convolutional neural network,the training of the model needs to select the data set with appropriate speckle distribution characteristics.Most of the existing SAR image data sets are synthetic data sets for machine learning model training.This paper proposes to use multi-temporal multi-look method to construct real SAR image data sets,which can effectively improve the speckle suppression ability of the network model.(2)In this paper,a multi-scale convolutional neural network(MSCNN)SAR image speckle suppression method with lightweight multi-scale interactive structure is proposed.This method extracts multi-scale features from SAR images through convolutional layers of different depths and stitches them to expand the receptive field of convolutional neural networks,which can extract local to global feature information of SAR images.The effectiveness of the network structure is effectively verified by simulated SAR image experiments.Through the comparative analysis of real SAR image experiments and many classical algorithms,it is proved that the proposed SAR image speckle suppression method MSCNN has better speckle suppression performance and efficiency.(3)A channel-attention multi-scale convolutional neural network(CAMSCNN)SAR image speckle suppression method is proposed.This method adds a channel attention mechanism to the multi-scale feature fusion module of the MSCNN method,which can dynamically extract multi-scale SAR image features,effectively overcome the data redundancy problem in the model data training process,and improve the model performance.Firstly,multiscale features are extracted from SAR images through convolution layers of different depths,and then stitched.Then,the channel attention mechanism is introduced to assign different weights to different scale features,and the weighted multi-scale fusion features are obtained.Finally,the final speckle suppression SAR image is generated by global residual noise reduction and image structure fine-tuning.In this paper,a variety of scenes of simulated SAR data and real SAR data are set up for speckle suppression experiments.The results show that the proposed CAMSCNN method performs best in quantitative evaluation indicators under different scenes and different noise views,and shows great advantages in visual effects,indicating that the proposed CAMSCNN method maintains a balance between speckle suppression intensity and detail retention of SAR images.The method proposed in this paper can provide key technical support for the practical application of SAR images.