IEEE Trans Image Process. 2016 May;25(5):2249-58. doi: 10.1109/TIP.2016.2545248.
For the hyperspectral target detection, the neighbors of a target pixel are very likely to be target pixels, and those of a background pixel are very likely to be background pixels. In order to utilize this spatial homogeneity or smoothness, based on total variation (TV), we propose a novel supervised target detection algorithm which uses a single target spectrum as the prior knowledge. TV can make the image smooth, and has been widely used in image denoising and restoration. The proposed algorithm uses TV to keep the spatial homogeneity or smoothness of the detection output. Meanwhile, a constraint is used to guarantee the spectral signature of the target unsuppressed. The final formulated detection model is an ℓ1-norm convex optimization problem. The split Bregman algorithm is used to solve our optimization problem, as it can solve the ℓ1-norm optimization problem efficiently. Two synthetic and two real hyperspectral images are used to do experiments. The experimental results demonstrate that the proposed algorithm outperforms the other algorithms for the experimental data sets. The experimental results also show that even when the target occupies only one pixel, the proposed algorithm can still obtain good results. This is because in such a case, the background is kept smooth, but at the same time, the algorithm allows for sharp edges in the detection output.
对于高光谱目标检测，目标像素的邻域很可能是目标像素，而背景像素的邻域很可能是背景像素。为了利用这种空间同质性或平滑性，基于全变差（Total Variation，TV），我们提出了一种新的基于单目标谱的有监督目标检测算法。TV 可以使图像平滑，已被广泛应用于图像去噪和恢复。所提出的算法利用 TV 保持检测输出的空间同质性或平滑性。同时，使用约束来保证目标的光谱特征不受抑制。最终的公式化检测模型是一个 l1 范数凸优化问题。分裂布格曼算法（Split Bregman Algorithm）用于解决我们的优化问题，因为它可以有效地解决 l1 范数优化问题。我们使用两个合成和两个真实的高光谱图像进行实验。实验结果表明，对于实验数据集，所提出的算法优于其他算法。实验结果还表明，即使目标只占据一个像素，该算法仍然可以获得良好的结果。这是因为在这种情况下，背景保持平滑，但同时，算法允许检测输出中的尖锐边缘。
