Hyperspectral images are nonlinear and have a strong inter-spectral correlation. It is easy to lose some information when using a linear method to transform the dimension of hyperspectral data. In this paper, the kernel function is introduced into the minimum noise fraction (MNF), and the kernel minimum noise fraction (KMNF) is proposed. The data is mapped to the high-dimensional feature space through nonlinear mapping, and the minimum noise separation components are extracted in the high-dimensional space. The hyperspectral images have a strong inter-spectral correlation and a spatial neighborhood correlation, and the weights of the two wavebands and the spatial neighborhood are used for multiple linear regression processing to accurately estimate the noise of hyperspectral data. The constrained energy minimization (CEM) method and the matched filter (MF) method are the more classical methods in hyperspectral target detection. The KMNF is applied to two classical target detection methods, and the airfield data from AVIRIS data are used to carry out the hyperspectral target detection experiments.The results show that KMNF can well highlight targets and improve the detection effect and accuracy of hyperspectral targets.