-
摘要: 高光谱异常检测因其以无监督方式检测目标的能力而受到特别关注,自动编码器及其变体可以自动提取深层特征,还可以检测异常目标。由于训练集中存在异常,自动编码器泛化性较强,从而降低了从背景中区分异常的能力。为解决上述问题,该文提出一种基于弱监督鲁棒性自编码的异常探测算法。首先提出了一种显著类别搜索策略,采用基于概率的类别阈值来标记粗样本,为网络弱监督学习做准备; 同时,构建一个具有l2,1范数与异常-背景光谱距离共同约束的鲁棒性自编码网络框架,该框架在训练期间对噪声和异常具有鲁棒性; 最后,采用所有样本得到的重构误差检测异常目标。在4个高光谱数据集上进行实验,结果表明,与其他先进的异常检测算法相比,所提算法具有更好的检测性能。Abstract: Hyperspectral anomaly detection has received particular attention due to its unsupervised detection of targets. Moreover, autoencoder (AE), together with its variants, can automatically extract deep features and detect anomalous targets. However, AE is highly generalizable due to the existence of anomalies in the training set, thus suffering a reduced ability to distinguish anomalies from the background. This study proposed an anomaly detection algorithm based on the weakly supervised robust AE (WSRAE). First, this study developed a salient category search strategy and used probability-based category thresholds to label coarse samples in order to make preparation for network-based weakly supervised learning. Moreover, this study constructed a robust AE framework constrained jointly by l2,1 norm and anomaly-background spectral distances. This framework was robust with regard to noise and anomalies during training. Finally, this study detected anomalous targets based on the reconstruction errors obtained from all samples. Experiments on four hyperspectral datasets show that the WSRAE algorithm has greater detection performance than other state-of-the-art anomaly detection algorithms.
-
Key words:
- hyperspectral image /
- anomaly detection /
- salient category search /
- robust AE
-
-
计量
- 文章访问数: 34
- PDF下载数: 5
- 施引文献: 0
下载: