| Summary: | Representation-based target detectors for hyperspectral imagery have attracted considerable attention in recent years. However, their detection performance is still unsatisfactory due to the independent manner of the recovery process on each test pixel. Moreover, the background dictionary generated through the dual windows is susceptible to target contamination. Aiming to address these issues, in this article, we propose a decomposition model (DM) with background dictionary learning (BDL) for hyperspectral target detection. The observed data are decomposed into three parts: background, target, and noise. The background and target dictionaries are utilized to represent the background and target components, respectively. In order to achieve a satisfactory recovery of the background and target components, the proposed DM exploits the spatial smoothness of background pixels and the scarcity of the targets of interest in the whole scene via the total variation and the sparsity, respectively. Then, the separated target image is directly used for the detection purpose. Furthermore, a novel BDL method based on the locality-constrained linear coding is presented, and a complete and compact background dictionary can be learned with a low computational cost. Meanwhile, the a priori target dictionary is also incorporated into the learning process in order to suppress the contamination of the target signal on the learned background spectra. Extensive experiments on both simulated and real hyperspectral datasets demonstrate the superiority of the proposed detector in comparison with several conventional and state-of-the-art target detectors.
|
|---|
