Anomaly Detection From Hyper Spectral Image

Due to the unique imaging characteristics, acquiring spectral and spatial information simultaneously, hyper spectral images provide us an effective method to target detection, especially in complex background. Recently, hyper spectral target detection has becoming a charming topic in automatic target detection. Anomaly detection, without using the apriori spectral information of the target, is performed by the spectrum difference between the tested pixel and those in its neighbors, so it is useful in practice. In this paper, several hyper spectral imagery anomaly detection methods were studied on the hyper spectral imagery characters. Generally speaking, anomaly detection is implemented by a two-stage process, first by anomaly detection to find potential targets, followed by target discrimination to achieve target classification at the second step.Firstly, to provide a theory foundation for creating anomaly detection arithmetic operators and simulation data, the spectral and spatial characteristics of the hyper spectral image were discussed .Secondly, according to the size the object, the existed anomaly detection methods are classified into two types: point target anomaly detection and area target anomaly detection. For point anomaly detection, several classical anomaly detection methods were discussed including RX detection and its modified versions, UTD detection, and so on. Based on the spatial characteristics of hyper spectral images, a new method based on the neighborhood correlation and the least square algorithm were proposed. The presented method improved anomaly detection ability in the complicated background dramatically. For area target anomaly detection, an improved method based on RX detection were presented. The main contributions of the improved method include that (1) enhancing target signals and improving the Gaussian specificity of the background by a residual image of original hyper spectral image and that estimated by the least square algorithm, and (2) a new method for automatic threshold segmentation is proposed.At last, a background inpainting method based on tensor voting technique is utilized to remove the false detection results provided the above methods (including point and area target detection methods). The targets were detected by spectrum matching technique finally.