The traditional spectral dimension reduction methods are usually carried out by matching the reconstructed spectra to the original spectra mathematically, which will often result in reconstructed spectra of small spectral reconstruction errors but very poor colorimetric accuracy when compared with the original one. In order to minimize both the spectral and colorimetric errors more efficiently, we proposed three spectral dimension reduction methods by introducing the characteristics of human vision. The first method is VPCA, in which we apply spectral luminous efficiency function to the original spectra before reduction; The Second method (LMSPCA) uses a matrix derived from LMS cone sensitivity to weight the original spectra before reduction, and the matrix can be form by two methods, in which the L, M, S cones response offset is calculated by in two different ways: one is computed as the absolute value of each corresponding wave length offset, and the other is calculated as the square of each corresponding wave length offset. The third method is LMSPCAs, which is based on the second method LMSPCA by further applying PCA to the residual spectra. The result shows that the VPCA method produces the poorest perfomance. The two cones response weighted matrixes of LMSPCA method have similar performances by presenting better colorimetric accuracy and low spectral accuracy, while LMSPCAs method which compensates for the spectral loss of LMSPCA method can produce higher spectral and colorimetric reconstruction accuracy and color stability under different light source, and satisfies the requirements of spectral color reproduction.