Flexible Background Modeling for Hyperspectral Anomaly Detection Using Convolutional Sparse Coding

Abstract

Translate Article

This paper proposes a hyperspectral (HS) anomaly detection method using a novel background model based on convolutional sparse coding (CSC). One promising HS anomaly detection approach is to simultaneously estimate background and anomaly components (and noise) from an HS image by solving an optimization problem that appropriately models each component. Existing methods are designed to capture the local pixel-wise relationships of a background component, but do not capture non-local similarity structures. As a result, these methods misclassify widely distributed and relatively large objects in a background component as a part of an anomaly component. Therefore, we propose improving the modeling of a background component by using CSC. Specifically, we formulate an optimization problem that incorporates HS total variation and CSC to capture local and non-local similarity structures, respectively. We then develop an algorithm to solve this problem based on alternating minimization. Experiments on HS datasets show that our method achieves better detection performance than existing state-of-the-art methods.