Cross-AD: Multispectral and Hyperspectral High-Speed Artificial Imitation Object Anomaly Detection

Multispectral and hyperspectral imaging systems have shown great potential in detecting artificial imitations at natural backgrounds. However, the general spectral anomaly detection(AD) algorithms lack attention to the camouflage characteristics and are not fully suitable for real-time inspection applications. Since the imitative targets have local extent similarity but global differences between real backgrounds, we proposed the Cross-AD method for imitation anomaly detection (IAD), which is based on the horizontal and vertical adaptive background estimation at the pixel level. It combines the scattered distribution features of camouflage and the directional model has a better suppression effect on the common band noise, while the execution time is close to the RX detector. Furthermore, the improved Cross-Box with protection window and Cross-Index with feature factor is proposed to better deal with large imitations and dense vegetation environments. To validate the algorithms, the hyperspectral imitation anomaly detection (HSIAD) dataset is constructed based on the actual camouflage scene. Compared with the real-time and fast spectral AD detection algorithms, the cross-series methods achieve the optimal balance of execution time and detection performance on both multi-and hyperspectral IAD datasets. The implementation code is available at https://github.com/XingshiLuo/Cross-AD.