结合扩张卷积的残差网络SAR图像去噪  

Residual network combined with dilated convolution for SAR image denoising

作　　者：申兴成 杨学志 董张玉[1,2] 陈鲸 SHEN Xingcheng;YANG Xuezhi;DONG Zhangyu;CHEN Jing(School of Computer and Information,Hefei University of Technology,Hefei 230601,China;Anhui Province Key Laboratory of Industry Safety and Emergency Technology,Hefei 230601,China;School of Software,Hefei University of Technology,Hefei 230601,China;Intelligent Interconnected Systems Laboratory of Anhui Province,Hefei 230601,China)

机构地区：[1]合肥工业大学计算机与信息学院,合肥230601 [2]工业安全与应急技术安徽省重点实验室,合肥230601 [3]合肥工业大学软件学院,合肥230601 [4]智能互联系统安徽省实验室,合肥230601

出　　处：《测绘科学》2021年第12期106-114,共9页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(41601452);安徽省科技攻关计划项目(202004a07020030)。

摘　　要：针对现今的合成孔径雷达图像去噪算法对图像去噪后结构保持不足的问题,该文提出了一种增强的残差卷积神经网络(ERCNN)。首先,ERCNN构建了增强的残差卷积模块,该模块通过结合卷积和扩张卷积,扩大网络的感受野;其次,通过加入残差连接将局部特征信息从低层传递到更高层,减少梯度消失问题;再次,使用批归一化加速模型的训练速度;最后通过残差学习的策略,形成端到端的映射以去除噪声。通过仿真和真实合成孔径雷达图像进行实验的结果表明,所提出的ERCNN具有优于现有方法的性能,在去噪的同时还可以保留更多的细节信息,且拥有高效的计算效率。In order to overcome the shortcoming of current SAR image denoising methods on preserving structural information,an enhanced residual convolutional neural network,named ERCNN,was proposed in this paper.First,enhanced residual convolutional neural network(ERCNN)constructed an enhanced residual convolution module,which enlarged the receptive field of the network by combining the convolution and the dilated convolution.Then skip connections were employed to prevent the loss of local feature information and reduced the vanishing gradient problem.In addition,batch normalization was used to accelerate the training speed of the model.Finally,an end-to-end mapping was formed to remove noise through a residual learning strategy.Experiments on simulation and real SAR images showed that the proposed ERCNN had better performance than state-of-the-art methods,more details could be preserved while denoising,and it also maintained high computational efficiency.

关 键 词：SAR图像 去噪 扩张卷积 残差连接 结构保持 残差学习 

分 类 号：P237[天文地球—摄影测量与遥感]