The multispectral image acquisition oriented to reproduction requests that the data is device independent and scenes independent, and can realize the characterization of the original color information. Aiming at disturbance, noise error of system, and the requirement for training samples’ typical representative and correlation, the authors proposed orthogonal regression spectral algorithm and training samples selection algorithm based on subspace tracking, through the mapping function between the spectral space and color space, by selecting the best samples in typical representative and correlation samples between target samples and selected samples. The modified Sinar 75H trichromatic digital camera combined with bandpass filter glasses were used for experiment, the data show that our method has higher spectral and chromaticity accuracy, the training samples selected by subspace tracking method are uniformly distributed in the sample space, and have good orthogonality. The statistics experimental results indicate that the performance of the proposed method is obviously better than that of previous method, in both color difference error and spectral reflectance error.