Bayesian Constrained Energy Minimization for Hyperspectral Target Detection

Abstract

For better performance of hyperspectral target detectors, the prior target spectrum is expected to be accurate and consistent with the target spectrum in the hyperspectral image to be detected. The existing hyperspectral target detection algorithms usually assume that the prior target spectrum is highly reliable. However, the label obtained is not always precise in practice, and pixels of the same object may have quite different spectra. Since it is hard to acquire a highly reliable prior target spectrum in some application scenarios, we propose a Bayesian constrained energy minimization (B-CEM) method for hyperspectral target detection. Instead of the point estimation of the target spectrum, we infer the posterior distribution of the true target spectrum based on the prior target spectrum. Specifically, considering the characteristics of hyperspectral image and target detection task, we adopt the Dirichlet distribution to approximate the true target spectrum. Experimental results on three datasets demonstrate the effectiveness of the proposed B-CEM when the known target spectrum is noisy or inconsistent with the true target spectrum. The necessity to approximate the true target spectrum is also proved. Generally, the distributional estimate achieves better performance than using the known target spectrum directly.