摘要
高光谱异常检测以无监督方式分析背景和异常目标的光谱特征,探寻异常目标的分布情况。受高光谱背景复杂分布的影响,训练样本中含有异常使得深度网络模型泛化性弱,并且难以做到直接预测异常目标,降低了模型的实际应用能力。提出一种基于自监督多尺度差分的异常检测算法,旨在提升模型的泛化能力,实现网络模型直接预测异常目标。首先设计基于测度K-means的显著类别搜索策略,结合熵信息测度与类间距离测度优化标记伪异常与背景样本,以自监督学习的方式迭代更新样本。此外,基于中心差分卷积构建多尺度差分网络结构,经过多尺度特征融合与概率预测,实现以端到端的方式检测异常目标。通过4个高光谱数据上实验,结果表明所提算法具有更好的检测性能,同时能更好地抑制背景的干扰。
Hyperspectral anomaly detection analyzes the spectral features of both background and anomalous targets in an unsupervised manner to explore their distribution.Due to the complex distribution of hyperspectral backgrounds,training samples may contain anomalies that weaken the deep network model’s generalization ability.Directly predicting abnormal targets is challenging for the network model,which reduces its practical application.In this paper,we proposed a multi-scale differential anomaly detection method based on self-supervised to improve the model’s generalization ability and enable it to predict anomalous targets directly.We designed a significant category search strategy based on K-means,and combined the entropy information measure and inter-class distance measure to optimize the marking of pseudo-anomalies and background samples.We iterated and updated the samples in a self-supervised learning manner.Additionally,we constructed a multi-scale differential network structure based on center differential convolution,and achieved the end-to-end detection of anomalous targets by multi-scale feature fusion and probability prediction.Experimental results of four hyperspectral data show that this algorithm has better detection performance while suppressing background interference effectively.
作者
周晓忠
刘军廷
周海涛
ZHOU Xiaozhong;LIU Junting;ZHOU Haitao(Ningbo Metallurgical Survey,Design and Research Co.,Ltd.,Ningbo 315194,China)
出处
《地理空间信息》
2025年第6期37-42,共6页
Geospatial Information
关键词
高光谱异常检测
K-MEANS聚类
中心差分卷积
多尺度
hyperspectral anomaly detection
K-means clustering
center differential convolution
multi-scale
