面向高光谱异常检测的背景记忆模型

A background memory model for hyperspectral anomaly detection

高光谱图像具有丰富的连续光谱信息,覆盖从可见光到红外波长范围内数百个波段。此数据特性使得高光谱数据在图像处理中有挖掘光谱内在特征的独特优势,从而有利于充分利用空间和光谱信息,检测感兴趣区域中的目标。但是由于高光谱数据的高维性、实际场景的复杂性以及有标签样本数量的有限性,高光谱异常检测仍面临背景和异常不易区分和检测精度低的问题。为了解决以上问题,本文提出一种面向高光谱异常检测的背景记忆模型。首先,通过无监督的基于密度估计的粗检方法,获得伪背景和伪异常向量。其次,设计基于异常突出正则项约束的背景记忆生成对抗网络模型,扩大伪背景和伪异常间的距离,并采用弱监督—伪标签的训练方式,使得网络具有较强的背景生成能力,同时弱化其对于异常的重构效果,减少对于背景和异常重构的泛化性,增强输出检测结果中背景和异常的差异和辨别度。此外,在特征域和图像域进行对抗学习,提高样本生成能力,更好地学习到输入样本的分布情况,提高网络对背景的生成能力。最后,使用非线性背景抑制方法减小虚警率,进一步提升检测精度。实验结果表明,本文方法相比其他检测算法在不同数据集上具有更好的检测效果。

Hyperspectral images(HSIs)have a wealth of continuous spectrum information,covering hundreds of bands from visible light to infrared wavelengths.The data characteristics of HSIs give it a unique advantage in harnessing the inherent attributes of the spectrum in image processing.This advantage is conducive to making full use of spatial and spectral information and detecting targets in the region of interest.However,due to the high dimensionality of hyperspectral data,the complexity of actual scenes,and the limited number of labeled samples,hyperspectral anomaly detection faces the problem of indistinct background and anomalies and low detection accuracy.Therefore,we propose a background memory model for hyperspectral anomaly detection.First,the pseudo background and anomaly vectors are obtained through an unsupervised rough inspection method based on density estimation.Second,we design a background memory generation adversarial network model based on anomalous prominent regular term constraints.Moreover,we expand the distance between the false background and false anomalies in a weak supervision-pseudo-label manner.Thus,the network has a strong background generation ability while the effect on anomaly reconstruction is weakened,which reduces the generalization of background and anomaly reconstruction and enhances the difference and discrimination between background and anomaly.We also perform adversarial learning in the feature domain and image domain to improve sample generation ability,enabling better learning of the distribution of input samples and strengthening the capability to generate background.Finally,a nonlinear background suppression method is introduced to reduce the false alarm rate and further improve the detection accuracy.The experimental results show that our model has a better detection effect on different datasets than other detection algorithms.HSIs have continuous spectral information of hundreds of bands,which make it possible to capture the deep and intrinsic characteristics in a spectrum.However,due to the high dimension of HSI,the complexity of the scene,and the limitation of labeled samples,hyperspectral anomaly detection remains a challenge.To solve the abovementioned problem,we propose a generative adversarial network with anomalyhighlighted regularization and train it in a weakly supervised manner.We aim to separate the anomaly and background vectors to make the difference more obvious and obtain a more accurate detection map.In this paper,we propose a background memory generative adversarial network for hyperspectral anomaly detection.First,we obtain the pseudo background and anomalies through unsupervised coarse detection based on density estimation as the input of the network.Next,to reduce anomaly contamination in background estimation,we impose the constraint of anomaly-highlighted regularization to expand the distance between the background and anomaly.In the weak supervised pseudo labeling training mode,the network can reconstruct background vectors well but gains poor performance for anomaly reconstruction.Besides,there are two discriminators in the latent and reconstruction domains,which aim to improve the ability of background generation and estimation.Finally,we perform nonlinear background suppression on the detection map as post-processing to reduce the false alarm rate.Compared with other new algorithms with good performance,the proposed method has better detection results in both quantitative and qualitative aspects of different datasets.The AUC score of(Pd,Pf)achieves the highest value across different datasets and outperforms other algorithms and has the advantage of an order of magnitude.For example,the AUC score of(Pd,Pf)achieves 0.99771 for the ABU-1 dataset,while the AUC score of(Pf,τ)is 0.00258,which outperforms the second-best algorithm AED with scores of 0.99760 and 0.02230,respectively.The visual results are consistent with the qualitative results as well.The ROC curve locates near the upper left corner.Under the same false alarm rate,the proposed method has the highest accuracy for most datasets,obtaining higher detection probability and lower false alarm rate,and has better detection performance.The box plot likewise reveals that the background and anomaly of this method are more separable.In this paper,we propose a generative adversarial network memorizing background for hyperspectral anomaly detection.Different from our previous work,we obtain the pseudo background and anomaly vector adaptively in an unsupervised manner to solve the problem of the small number of anomaly samples and lack of prior information.Based on weakly supervised pseudo-labeling learning,we aim to model a hyperspectral anomaly and background vector.As a result,the network can reconstruct background data well but performs poorly on anomaly vector reconstruction.We also apply the constraint of highlighting an anomaly regular term in the network to enhance the separability between background and anomaly.Finally,we perform post-processing of nonlinear background suppression to reduce the false alarm rate under the same detection accuracy.Experimental results show that the proposed method can achieve better detection performance than other algorithms on different datasets.