Hyperspectral Anomaly Detection Based on Intrinsic Image Decomposition and Background Subtraction

Hyperspectral anomaly detection is a detection of abnormal targets in a region based on spectral and spatial information under the premise of no prior knowledge of the target, which is a very important research topic in the field of remote sensing. In the anomaly detection of hyperspectral images, the salient feature map and spatial-spectral features are not effectively used, which greatly limits the detection performance. To solve the above problems, this paper proposes a hyperspectral anomaly detection method based on intrinsic image decomposition and background subtraction. Firstly, the optimal clustering framework is used to select the appropriate bands as the subsequent input images. Secondly, the hyperspectral visual attention model is applied to extract the salient feature map of the image, and the initial anomaly detection map can be obtained by morphological filtering and background subtraction. Then, the pure spectral information in the reflection component of the hyperspectral image is obtained by the intrinsic image decomposition, and the adaptive weight map is calculated on the reflectance image by the spectral angle distance. Finally, the weight map is fused with the initial anomaly detection map to obtain the final anomaly detection result. In the experimental, the proposed method is compared with eleven anomaly detection methods, which including GRXD, RPCA-RX, LSMAD, LRASR, GTVLRR, PTA, HAD-LEBSR, TPCA, PCA-TLRSR, VABS, and GNLTR. The results demonstrate that the proposed method can better enhance the separability of background and anomaly target, so it improves the accuracy of anomaly detection.