边缘信息引导学习的改进SAR-CNN相干斑抑制算法  

Improved SAR-CNN suppression algorithm of speckle with edge information

作　　者：樊雯雪 朱磊[1] 朱奇伟 李志蒙 姚同钰 FAN Wenxue;ZHU Lei;ZHU Qiwei;LI Zhimeng;YAO Tongyu(School of Electronics and Information,Xi an Polytechnic University,Xi'an 710048,China;Hangzhou Shengqing Technology Co.,Ltd,Hangzhou 310052,Zhe Jiang,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048 [2]杭州昇擎科技有限公司,浙江杭州310018

出　　处：《长江信息通信》2022年第11期41-45,共5页Changjiang Information & Communications

基　　金：陕西省重点研发计划(2019GY-113);陕西省自然科学基础研究计划(2019JQ-361)。

摘　　要：为提升传统CNN网络在抑制SAR图像乘性相干斑时的边缘保护性能,提出了一种改进的CNN抑斑算法SARICNN。该方法首先利用一阶均值比与二阶均值比联合构建的图像强度信息,改善受乘性相干斑影响的边缘检测性能,并经阈值化处理获取图像的边缘区域与同质区;其次,对不同区域给予不同的权重,生成突出边缘信息的边缘强调图,并将边缘强调图与含斑图像一起送入CNN卷积神经网络,引导网络更精确地学习图像边缘信息,从而在抑斑的同时更好保护边缘。实验结果表明,与SAR-CNN等五种算法相比,SAR-ICNN算法获得的抑斑图像具有更清晰的边缘视觉效果与更高的参数指标,其中对Set12数据集形成的仿真SAR图像,PNSR、SSIM、EPI三类指标分别平均提升了1.30%、1.16%、4.51%;对真实SAR图像,ENL、EPI两类指标分别提升了116.93%与23.24%。In order to improve the edge protection performance of traditional CNN network in suppressing multiplicative coherent speckle in SAR images,an improved CNN speckle suppression algorithm SAR-ICNN was proposed.The method firstly uses the image intensity information jointly constructed by the first-order mean ratio and the second-order mean ratio to improve the edge detection performance affected by the multiplicative coherent speckle,and obtains the edge area and homogeneous area of the image through thresholding;Different areas are given different weights to generate an edge-emphasized map that highlights edge information,and the edge-emphasized map is sent to the CNN convolutional neural network together with the speckled image to guide the network to learn the image edge information more accurately,so as to suppress speckle at the same time.Better edge protection.The experimental results show that,compared with five algorithms such as SAR-CNN,the speckle suppression image obtained by the SAR-ICNN algorithm has a clearer edge visual effect and higher parameter indicators.The simulated SAR image formed by the Set12 dataset,PNSR,SSIM,and EPI indexes increased by 1.30%,1.16%,and 4.51%respectively;for real SAR images,ENL and EPI indexes increased by 116.93%and 23.24%on average.

关 键 词：SAR图像 相干斑抑制 卷积神经网络 边缘引导 边缘保持 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]