高光谱图像(HyperSpectral Images,HSIs)具有高光谱分辨率和丰富的光谱信息,其具有的大量窄波段电磁波有利于获取感兴趣目标的理化信息,并根据对应的光谱特征对不同物质进行有效区分,从而完成目标检测任务.然而有限样本、少量先验信息、高维相似背景及不同类别差异小所导致的目标和背景混淆问题使得高光谱目标检测(Hyperspectral Target Detection,HTD)面临挑战.为此,本文提出结合区域引导和双注意力机制的高光谱目标检测判别式学习网络(Region-guided and dual-Attention Discriminative learning Network,RADN),以缓解标记样本少的条件下不同类别相似度高和相同类别差异性大导致的背景和目标不易区分的问题,减少高维冗余特征带来的计算复杂度,同时提升检测精度.本文使用经验性区域引导网络训练,采用光谱约束的无监督聚类方法确定网络输入,选择性地关注高光谱图像中的显著性特征和感兴趣区域.此外,本文在网络中添加双通道注意力机制来辅助复杂背景分布的估计,并在网络中引入不同类别光谱先验损失函数,进一步减少高维复杂背景以及光谱变化对于目标的干扰.实验结果和分析表明,RADN在不同数据集上的性能优于现有先进的算法.
Region-Guided and Dual Attention Discriminative Learning Network for Hyperspectral Target Detection
Hyperspectral images(HSIs)have high spectral resolution and rich spectral information,which can obtain the physical and chemical information of the target of interest by using a large number of narrow-band waves.HSIs can ef-fectively distinguish different substances by corresponding spectral features,and complete the task of target detection.How-ever,the problem of target and background confusion caused by limited samples,a small amount of prior information,high dimensional similar background,and differences between different classes make hyperspectral target detection(HTD)still face challenges.To this end,we propose a region-guided and dual-attention discriminative learning network(RADN)for HTD to solve the problem of intra-class differences and inter-class similarities under a few samples.It can reduce the com-putational complexity caused by high-dimensional redundant features and improve detection accuracy.In this paper,we in-troduce the empirical region-guided network for training.We employ the spectrally constrained unsupervised clustering net-work to determine the network input.To selectively focus on salient features and regions of interest,we add a dual-channel attention mechanism in the generator and discriminator to assist in the estimation of complex background distributions;We introduce an inter-class spectral prior loss function in the network and further reduce the interference of high-dimensional complex background and spectral changes to the target.Experimental results and analysis show that RADN outperforms ex-isting state-of-the-art algorithms on different datasets.
hyperspectral target detectionunsupervised clusteringchannel attention mechanismregion of interestspectral resolution
国家科技期刊平台
NSTL
万方数据
