Hyperspectral Anomaly Detection via Anchor Generation

Hyperspectral anomaly detection stands out as a focal point within the realm of hyperspectral image (HSI) analysis. However, traditional anomaly detectors face challenges due to the high susceptibility of background pixels to anomaly contamination. Moreover, they often fall short in adequately leveraging the spectral and spatial information present in HSIs. In response, we propose a hyperspectral anomaly detection method employing Anchor Generation, which effectively addresses the aforementioned issues by enhancing the distinction between anomalous and background pixels. Firstly, we sort pixels by computing the Mahalanobis distance (MD) between them and the background, thereby acquiring information on the degree of anomalies. This aids in distinguishing abnormal pixels from background pixels. Secondly, we enhance background purity by merging pixels close to anomaly-free regions. Moreover, we design a local MD (LMD) algorithm to better extract spectral and spatial information from the HSI, facilitating better reflection of local features. Through the integration of anomaly detection results from global and two local regions using logicalor andand operations, we enhance the accuracy and robustness of anomaly detection.