为滤除大量冗余背景信息, 提升目标检测速度, 解决目标光谱获取困难等问题, 提出一种融合光谱微分重排与光谱角匹配的高光谱快速目标检测算法。首先, 对已知背景光谱信息微分重排, 筛选出背景特征较少的谱段集; 然后, 计算相应谱段下, 高光谱图像各像素点的光谱微分值与背景光谱微分值的差值, 并对所得各谱段微分差值加权求和阈值化, 粗提取出目标位置; 最后, 计算提取出的目标位置光谱与先验背景光谱角匹配的反余弦值, 实现目标背景差异精细判别。通过自行拍摄的草地上黄、绿伪装网及AVIRIS获取的San Diego机场飞机高光谱图像进行实验, 与RX、CEM、OSP、ACE、SAM等算法比较。实验结果表明, 所提算法在保证检测精度的同时, 具有较低的虚警率和较好的时效性: ROC曲线下面积AUC均高于0.98, 虚警率分别为3%和1.6%, 处理时间仅为0.36 s和0.077 s。

In order to filter a large amount of redundant background information, improve the speed of target detection and solve the difficult access problem of target spectrum, a fast hyperspectral target detection algorithm which is fused on spectral derivative, recomposition and spectral angle mapper is proposed. Firstly, the derivative of background spectrum is calculated and reordered to find a spectral set with less background information. Secondly, each pixel of the spectral image in the corresponding spectral set are used to calculate the spectral derivative results subtracting the spectral derivative results of background, then the threshold derivative difference of each spectrum is weighted and summed to extract the target location roughly. Finally, spectral angle mapper algorithm is utilized to calculate the arccosine values of target points and priori background to distinguish the difference between targets and background precisely. Two hyperspectral images of yellow and green camouflage nets on the grass and the planes on the San Diego airport captured by AVIRIS were selected in the experiment for comparing the proposed algorithm with RX, CEM, OSP, ACE, SAM algorithms. The experimental results indicate that the proposed algorithm has a good detection precision, low false alarm rate and high efficiency with AUC higher than 0.98, false alarm rate 3% and 1.6% respectively, the processing time only 0.36 s and 0.077 s.