SAR Image Despeckling Based On Deep Learning

Synthetic Aperture Radar(SAR)is an active coherent imaging radar that can image the target area all day and under all weather conditions,providing an abundant information supplement for optical remote sensing images.But its coherent imaging method brings the SAR image with strong and random speckle,which seriously affects subsequent applications of SAR images.Therefore,it is of great practical significance to study the SAR image despeckling.In order to obtain high-quality despeckling results,the traditional SAR image despeckling algorithms often introduce complex optimization problems,which not only have a large amount of computation and consume a lot of time,but also have many hyperparameters that require some experience to set.In contrast,SAR image despeckling algorithms based on deep learning can not only automatically extract features of the image from amounts of data,but also can accelerate the network forward propagation process through the GPU,so both despecking performance and efficiency are improved compared to the traditional algorithms.Therefore,in this paper,we choose to study the SAR image depeckling algorithm based on deep learning theory,the innovative work of this paper mainly includes the following two parts:A SAR image despeckling network architecture based on simplified dense connections is proposed.The existing SAR image despeckling network architecture is generally simple,without effective use of the texture features extracted from the shallow layers of the network,and has some limitations in residual connection.In order to obtain a network with better despecking performance,first of all,this paper introduces dense connections and implements feature reuse between local network layers to form dense blocks;secondly,considering the computational efficiency of the network,a pattern pf simplified dense connection is proposed to reduce unnecessary connections in dense blocks;then,the simplified dense connections are added among dense blocks to realize feature reuse of the entire network,so as to effectively utilize the texture features extracted by shallow layers of the network;last but not least,by analyzing the speckle noise model,adding global and local residuals in the network to get the final network architecture.Based on the network architecture designed in the first part,in order to further improve the despeckling performance of the network,a SAR image despeckling algorithm combining deep learning and wavelet transform is proposed.Wavelet transform has excellent timefrequency analysis capabilities.It is possible to simultaneously concentrate the speckle noise and texture details of the SAR image in the high frequency subband.Therefore,we use wavelet coefficients instead of the original image as the input of the network.On the one hand,by setting the loss function with different constraint capabilities for different subbands,the network is guided to focus on the restoration of high-frequency coefficients,which achieves suppressing speckle noise effectively while preserving the texture details of the image,thereby the despeckling performance of the network are improved;on the other hand,Using the wavelet coefficient as input of the network can reduce the size of the feature maps,which could reduce the amount of calculation and speed up the network processing effectively.The experimental results show that the proposed SAR image desopeckling algorithm has excellent performance in terms of despecking performance,image texture structure preservation and processing efficiency.