Traditional principal component analysis (PCA) keep the similar shape of original spectral reflectance and reconstructing spectral reflectance as far as possible through mathematical method. But traditional dimension-reduction algorithm of PCA calculates and processes the spectral data each wavelength with equal weighted, while the sensitivity of human vision is different at different wavelength. It would result in that the spectral errors of reconstruction are small but the color differences of reconstruction color are large by human perception. In order to control the spectral error and reduce the chromatic difference between the original spectral and reconstructed spectral, this paper presents two kinds of human-vision-weighted function to optimize the traditional PCA, and using spectral residual error to compensate dimension-reduction model. With the experiment of training samples of Munsell color, and testing samples of multispectral image (young girl) and part of Munsell color, we reduced and reconstructed the spectral color and spectral image with our proposed-function-PCA, and compared with the other methods mentioned by related articles. The experimental results indicate that the performance of our methods improve the chromatic precision and stability in the different lighting resource.