Hyperspectral Image Anomaly Targets Detection with Online Deep Learning

Hyperspectral image (HSI) anomaly targets detection has been widely used in disaster alarm and military applications. Deep learning based HSI anomaly detector (AD) performs better by learning high-level features. However, the issues from heavy training computational burden and the model mismatch bring new challenges for online applications in the aspect of processing speed and detection accuracy. In this paper, an online Maximum-Distance-Pixel-Library(MDPL) method is proposed by using the most effective pixels to update deep auto-encoder based HSI AD with less extra computation. Experimental results on two recorded hyperspectral images show that the proposed method outperforms the traditional real-time local Reed-Xiaoli based detector in term of accuracy and processing time. Compared with fully updating deep learning based HSI AD, the proposed method performs higher time efficiency without accuracy loss.