| Synthetic Aperture Radar(SAR)with the ability of all-weather and all-time can access high-resolution surface images.However,speckle noise impairs the application of SAR images.Despeckling of SAR images becomes a hot research topic for remote sensing images.Due to limitations of traditional denoising methods,it is difficult to keep a good balance between the ability of despeckling and structure preservation.With the development of deep learning,convolutional neural network(CNN)has become a popular tool to encourage the development of SAR image despeckling.However,CNNbased despeckling methods cannot effectively preserve the image structure that need to be further improved.Inspired by recent work in SAR despeckling,two despeckling algorithms based on residual network have been proposed.The main research work of the thesis is as follows:(1)A SAR image despeckling algorithm is proposed through an enhanced residual convolutional network which combined with dilated convolution.Aiming at the problem that the existing CNN methods cannot preserve structure of the images effectively after denoising,firstly,an enhanced residual convolution module is constructed in this study that expands the receptive field by combining convolution and dilated convolution.Secondly,the information of local feature can transfer to deep layers from previous layers by adding skip connection;this process can use low-level features sufficiently and therefore reduce the vanishing gradient problem.Then,batch normalization is used to accelerate the speed of training the network.Finally,an end-to-end mapping is formed to remove speckle noise.(2)A SAR image despeckling algorithm is proposed based on the residual attention network and joint loss function.Aiming at the problem of weak correlation between feature maps in CNN,firstly,a residual convolutional block attention module is constructed to use channel attention to rebuild dependencies between channels,use spatial attention to strengthen the connection with global similar areas,and use the skip connection to join the input of module and the obtained output.Besides,feature fusion module is constructed to improves the capability of further feature extraction by splicing the optimized feature in channel dimension.Finally,the joint loss function is used to constrain the output of the network and improve the ability of images in despeckling and structure preservation.In order to verify the efficiency of the proposed algorithm,SAR data simulated by the UC Merced remote sensing image data set and the sentinel-1 SAR data are selected for experiments.Compared with the traditional methods and the state-of-the-art deep learning methods,experimental results demonstrate that the proposed algorithms have better results in visual effects and image quality metrics,and excellent denoising performance and structure preservation ability with efficient processing speed is shown. |
