An Edge Filtering-Based Spatial-Spectral Joint Hyperspectral Target-Level Anomaly Detection

Hyperspectral images have demonstrated exceptional performance in anomaly detection due to the strong distinctiveness of the spectral information they contain across different types of surfaces, drawing significant attention in applications such as civilian rescue operations and military search missions. However, commonly used hyperspectral anomaly detection techniques currently suffer from two major drawbacks: 1) existing anomaly detection algorithms overly focus on the spectral features of hyperspectral data while neglecting the spatial features contained within the image; 2) current detection algorithms typically operate at the pixel level, lacking relevant research aimed at achieving precise target set annotation. To address these two issues, this paper proposes an edge and keypoint detection-based hyperspectral image target set anomaly detection algorithm. First, a hyperspectral image edge enhancement operator based on the Scharr operator is designed to highlight target edge information by calculating the spectral similarity between pixels. Next, a spectral boundary-keypoint generation algorithm is proposed, which determines the coordinates of the target edge extrema by detecting edges in four directions. Finally, a target box generation algorithm is developed, combining boundary keypoints with anomaly detection results, and the results are optimized using Non-Maximum Suppression (NMS) by traversing combinations of extreme points. Extensive qualitative and quantitative experiments demonstrate that the proposed framework significantly improves the Intersection over Union (IoU) between the generated target boxes and the ground truth boxes while reducing the center offset, compared to state-of-the-art methods.