SAR图像目标轮廓增强预处理模块设计  

Design of SAR image target contour enhancement preprocessing module

作　　者：龚峻扬 付卫红[1] 刘乃安[1] GONG Junyang;FU Weihong;LIU Naian(School of Telecommunications Engineering,Xidian University,Xi’an 710071,China)

机构地区：[1]西安电子科技大学通信工程学院,陕西西安710071

出　　处：《系统工程与电子技术》2024年第12期4010-4017,共8页Systems Engineering and Electronics

摘　　要：针对合成孔径雷达(synthetic aperture radar,SAR)图像中的3个通道数据均相同、以至于在使用基于深度学习的目标检测网络对其进行目标检测时会造成通道信息利用效率低下的问题,基于对SAR图像中各种目标物轮廓特点的研究,提出一种基于平滑和锐化滤波的通道扩展预处理算法模块,并将其命名为ORLM(Original,Roberts,Laplace,Mean)模块。所提算法可被封装、集成并应用于目标检测算法的数据读取程序,可将每个通道数据均相同的SAR图像进行通道扩展,并且保证扩展后的通道数据充分包含目标的全部轮廓信息。通过将有否搭配本预处理算法的目标检测网络在不同舰船目标检测数据集上进行训练、测试及对比实验,实验结果表明,所提预处理算法可被应用在各种目标检测算法中,起到提升检测准确性的作用,且不会明显降低检测实时性。In response to the issue that the every channel’s data in three channels of synthetic aperture radar(SAR)images is the same which may cause channel information redundancy when deep learning-based target detection network detects the targets of SAR,a channel expansion preprocessing algorithm module based on smoothing and sharpening filtering is proposed,which is then named as ORLM(Original,Roberts,Laplace,Mean)block.The proposed algorithm in this article can be encapsulated,integrated,and applied to the data reading program of the target detection algorithm.It can extend the channel of SAR images with the same data in each channel,and ensure that the expanded channel data fully contains the contour information of the target.Through training,testing and comparative experiments of the target detection network with and without the proposed preprocessing algorithm on different ship target detection datasets,the experimental results show that the preprocessing algorithm proposed can be applied to various target detection algorithms and can improve detection accuracy without significantly reducing of the real-time detection performance.

关 键 词：图像处理 合成孔径雷达 深度学习 目标检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]