结合注意力机制的自监督高光谱图像异常检测  

Anomaly detection in hyperspectral images combining self-supervision and attention mechanism

作　　者：陈婷婷[1] 胡兴 刘德权 蒋林华 张大伟[1] CHEN Tingting;HU Xing;LIU Dequan;JIANG Linhua;ZHANG Dawei(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Yunnan Provincial Academy of Science and Technology,Kunming 650228,China;North China Institute of Science and Technology,Langfang 065201,China;Institut supérieur d'électronique de Paris-Sorbonne Joint Research Lab,Paris 92130,France)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093 [2]云南省科学技术院,昆明650228 [3]华北科技学院,廊坊065201 [4]巴黎高等电子学院-索邦联合实验室,巴黎92130

出　　处：《上海理工大学学报》2025年第1期45-53,共9页Journal of University of Shanghai For Science and Technology

基　　金：国家自然科学基金资助项目(62175037);云南省科技厅科技人才与平台计划(202305AF150143)。

摘　　要：高光谱图像异常检测广泛应用于农业、军事、地质、生物等领域。针对高光谱异常检测中数据样本少和空谱特性利用不足的问题,提出了一种结合注意力机制的自监督高光谱图像异常检测算法。首先,通过2D卷积自监督网络提取高光谱图像的光谱特征和空间特征来重构背景,引入注意力机制自适应的学习特征通道权重,实现特征优化;其次,考虑背景重构后图像信息量的减少问题,在损失函数中引入图像信息熵对特征编码进行约束,改善网络性能;最后,使用马氏距离实现异常值计算。将所提算法在两组来自不同场景的高光谱图像数据集上进行实验,并与7种同类算法进行了对比。结果表明,该算法在检测结果的AUC指标上均取得了最高值。Hyperspectral image anomaly detection is widely used in agriculture,military,geology,biology,and other fields.To solve the problem of few samples and insufficient utilization of space spectrum characteristics,an anomaly detection algorithm for hyperspectral images combining selfsupervision and attention mechanism was proposed.Firstly,a 2D convolutional self-supervised network was used to extract the spectral and spatial features of hyperspectral images to reconstruct the background.The attention mechanism adaptive learning feature channel weights were introduced in the network to optimize the feature.In addition,considering the reduction of image information after background reconstruction,image information entropy was added to the loss function to restrict the feature coding and improve the network performance.Finally,the anomaly value was calculated by the Mahalanobis distance.The proposed algorithm was experimented on two datasets of hyperspectral images from different scenes and compared with seven similar algorithms.The results show that the proposed algorithm has the highest value on area under curve(AUC).

关 键 词：高光谱图像 异常检测 自监督网络 注意力机制 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]