基于波段选择和空-谱组合核函数的高光谱图像目标检测

Hyperspectral image object detection based on band selection and spatial-spectral composite kernel

为了实现高光谱图像中特定目标的自动检测,提出了一种结合波段选择和空间-光谱特征组合核函数的高光谱图像目标检测方法。各像素点的光谱特征信息由所有波段构成的光谱曲线进行描述,空间特征信息则在各像素点周围的环绕局部范围内,采用词袋模型对空间灰度的纹理特征进行描述。使用了一种基于排序聚类的方法对波段进行选择以降低空间特征计算复杂度。模型训练阶段中,空间特征和光谱特征使用加权的形式融合为一个混合特征核,采用组合核函数结合支持向量机的方法优化核加权系数和检测模型的其他参数。实验结果表明,该方法将目标检测召回率提高到99.5%以上,虚警率降低到约0.2%。因此所提出的方法在降低波段数量的前提下,同时综合利用了目标的光谱信息和空间信息,并使空-谱两类特征在各类别上表现出重要性差异。

In order to automatically detect objects in hyperspectral images, we propose a new method in this paper for object detection in hyperspectral images based on combining band selection with spatial-spectral composite kernel. Spectral information for each pixel is represented by a spectral curve over all bands. Spatial information is represented by a bag of visual words model within a surrounding local region around each pixel. In order to reduce the computation complexity of spatial feature, a cluster-based band selection method is used before spatial feature extraction. In model training step, spatial and spectral features are fused into a composite kernel using weighted summation. Kernel weights and other detection model parameters are learnt under a framework combined a composite kernel and Support Vector Machine. Classification results demonstrate that the proposed method can yield a higher recall rate above 99. 5 % and a lower false alarm rate around 0. 2%. Therefore, the proposed method reduces the number of bands and utilizes spectral and spatial information of bands, while explaining the different importance of spectral and spatial features.