基于无监督图互信息最大化的海面小目标异常检测  

Anomaly Detection of Small Targets on Sea Surface Based on Deep Graph Infomax

作　　者：许述文[1] 何绮 茹宏涛 XU Shuwen;HE Qi;RU Hongtao(National Laboratory of Radar Signal Processing,Xidian University,Xi’an 710071,China)

机构地区：[1]西安电子科技大学雷达信号处理国家重点实验室,西安710071

出　　处：《电子与信息学报》2024年第7期2712-2720,共9页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61871303,62071346);高等学校学科创新引智计划(111计划)(B18039)。

摘　　要：受到复杂海洋环境的影响,雷达对海面慢速小目标难以实现高性能检测。对于这类目标,传统的基于能量的统计检测方法存在着严重的性能损失。针对这一问题,该文提出了基于互信息最大化框架下的海面小目标检测方法,实现海杂波背景下的无监督目标异常检测任务。首先,考虑到高分辨雷达回波不满足传统神经网络对样本独立同分布的假设,该文从图的角度重新建模数据,利用回波的空时相关特性来构建图拓扑结构。该文提出相对最大节点度并联合7个已有特征作为节点的初始表示向量。接下来,采用图注意力网络作为互信息最大化框架中的编码器学习节点表示向量。最后,使用异常检测算法进行目标检测,并实现虚警可控。经实测数据验证,使用快速凸包学习算法时,相比三特征检测器,所提检测器性能提升了9.2%;相比时频三特征检测器,性能提升了7.9%。当网络输出更高维的表示向量时,使用孤立森林算法的检测器的性能提升了27.4%。Due to the complex marine environment,it is difficult for a maritime radar to achieve high-performance detection of slow and small targets on the sea surface.For such targets,the traditional energy-based statistical detection algorithms suffer from serious performance loss.Confronted with this problem,a detection algorithm of small targets based on Deep Graph Infomax framework is proposed to realize unsupervised target anomaly detection in the background of sea clutter.In the traditional neural networks,there is an assumption that the samples are independent and identically distributed,which,however,the high-resolution radar echo does not meet.Therefore,this paper re-models the data from the perspective of graph and constructs the graph topological structure according to the correlation characteristics of the echo.Moreover,this paper puts forward the relative maximum node degree,and combines it with the relative average amplitude and the relative Doppler vector entropy to be the initial representation vectors of the graph nodes.With the graph modeling done,the graph attention network is used as the encoder in the Deep Graph Infomax framework to learn representation vectors.Finally,the anomaly detection algorithm is used to detect the targets,and the false alarm can be controlled.The detection result on the measured datasets shows that the performance of the proposed detector is improved by 9.2%compared to the three-feature detector when using the fast convex hull learning algorithm.Compared to the time-frequency three-feature detector,the performance is improved by 7.9%.When the network outputs a higher-dimensional representation vectors,the performance of the detector using the isolated forest algorithm is improved by 27.4%.

关 键 词：海杂波 目标检测 图神经网络 无监督学习 

分 类 号：TN957.51[电子电信—信号与信息处理]