刘畅,
1.中国科学院空天信息创新研究院 北京 100190
2.中国科学院电子学研究所 北京 100190
3.中国科学院大学 北京 100049
基金项目:国家重点研发计划(2017YFB0503001)
详细信息
作者简介:李萌：李 萌(1994–)，女，甘肃庆阳人，中国科学院大学硕士研究生，研究方向为SAR图像处理、机器学习。E-mail: limeng173@mails.ucas.ac.cn
刘畅：刘 畅(1978–)，男，山东烟台人，研究员，博士生导师。2006年在中国科学院电子学研究所获得博士学位，现担任中国科学院空天信息创新研究院研究员、博士生导师。主要研究方向为SAR系统及其相关SAR成像处理。E-mail: cliu@mail.ie.ac.cn
通讯作者:刘畅 cliu@mail.ie.ac.cn
责任主编：李宁 Corresponding Editor: LI Ning中图分类号:TN958
计量
文章访问数:2307
HTML全文浏览量:759
PDF下载量:328
被引次数:0
出版历程
收稿日期:2019-12-06
修回日期:2020-03-05
网络出版日期:2020-03-29
Super-resolution Reconstruction of SAR Images Based on Feature Reuse Dilated-Residual Convolutional Neural Networks
LI Meng,LIU Chang,
1. Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China
2. Institute of Electronics, Chinese Academy of Sciences, Beijing 100190, China
3. University of Chinese Academy of Sciences, Beijing 100049, China
Funds:The State Key Research Development Program of China (2017YFB0503001)
More Information
Corresponding author:LIU Chang, cliu@mail.ie.ac.cn
摘要
摘要:对于合成孔径雷达(SAR)图像，传统的超分辨重建方法对视觉特征的人为构造十分依赖，基于普通卷积神经网络(CNN)的超分辨重建方法对微小目标的重建能力较弱，对边缘轮廓的保真度较差。针对以上问题，该文提出一种基于特征复用的膨胀-残差卷积超分辨网络模型，同时引入感知损失，实现了精确的SAR图像4倍语义级超分辨。该方法为增加网络感受野，采用膨胀-残差卷积(DR-CNN)结构用于限制模型中特征图分辨率的严重损失，提高网络对微小细节的敏感度；为实现不同层级的特征最大化利用，将不同层级的特征图进行级联，形成一种特征复用结构(FRDR-CNN)，以此大幅度提升特征提取模块的效率，进一步提升超分辨精度；针对SAR图像特殊的相干斑噪声干扰，引入感知损失，使得该方法在恢复图像边缘和精细的纹理信息方面具有优越表现。文中实验表明，与传统算法以及目前较为流行的几种全卷积神经网络超分辨重建算法相比，该文采用的FRDR-CNN模型在视觉上对小物体的超分辨重建能力更强，对边界等轮廓信息的重建更准确，客观指标中的峰值信噪比(PSNR)和结构相似性指数(SSIM)分别为33.5023 dB和0.5127，边缘保持系数(EPD-ROA)在水平和垂直方向上分别为0.4243和0.4373。
关键词:合成孔径雷达/
超分辨重建/
膨胀卷积/
特征复用/
感知损失
Abstract:For Synthetic Aperture Radar (SAR) images, traditional super-resolution methods heavily rely on the artificial design of visual features, and super-reconstruction algorithms based on general Convolutional Neural Network (CNN) have poor fidelity to the target edge contour and weak reconstruction ability to small targets. Aiming at the above problems, in this paper, a Dilated-Resnet CNN (DR-CNN) super-resolution model based on feature reuse, i.e., Feature Reuse Dilated-Resnet CNN (FRDR-CNN), is proposed and perceptual loss is introduced, which accurately realizes four times the semantic super-resolution of SAR images. To increase the receptive field, a DR-CNN structure is used to limit the serious loss of the feature map’s resolution in the model, improving the sensitivity to tiny details. To maximize the utilization of features at different levels, the FRDR-CNN cascades the feature maps of different levels, which greatly improves the efficiency of the feature extraction module and further improves the super-resolution accuracy. With the introduction of the perceptual loss, this method has a superior performance in recovering image texture and edge information. Experimental results of the study show that the FRDR-CNN algorithm is more capable of providing small objects’ super-resolution and more accurate in the visual reconstruction of contour details, compared with traditional algorithms and several popular CNN super-resolution algorithms. Objectively, the Peak Signal to Noise Ratio (PSNR) is 33.5023 dB and Structural Similarity Index (SSIM) is 0.5127, and the Edge Preservation Degreebased on the Ratio Of Average (EPD-ROA) is 0.4243 and 0.4373 in the horizontal and vertical directions, respectively.
Key words:SAR/
Super resolution construction/
Dilated convolution/
Feature reuse/
Perceptual loss
PDF全文下载地址:
https://plugin.sowise.cn/viewpdf/198_a57f497d-d43c-4851-b2a6-811d1bf7eda0_R19110
