Tuning to Real for Single-Spectrum Hyperspectral Target Detection.

In many practical hyperspectral target detection, only a single spectrum of the target can be obtained before detection, while the original target spectrum is varied by the imaging condition. This severely degrades the practical accuracy and stability of existing target detection algorithms. To address this issue, we propose a novel tuning based approach to hyperspectral target detection with a single spectrum. In this approach, the given single target atom will be progressively tuned towards the real ones in the hyperspectral image; meanwhile, a pure background dictionary is optimized apart from the target atom to ensure their dissimilarity. Furthermore, we theoretically derive the sparse coefficient matrix of the target to assure a detector with lower false alarm rate and higher accuracy. Comparison results with related state-of-the-art methods on various datasets demonstrate that, our approach achieves the best detection performance in terms of accuracy and stability, when only a single spectrum of target is available.