Based on the characteristics of the color scanner, the spectral reconstruction of images was studied using the method combining the principal component analysis (PCA) and back-propagation (BP) artificial neural network. The IT8.7/2 standard color card was adopted as the training samples, and the different set of patches in this card was used as testing samples to discuss the effect of different structure networks, principal component numbers and training sample numbers on the spectral reconstruction; meanwhile the color patches of Natural Color System (NCS) were selected as testing samples to analyze the relationship between the performance of spectral reconstruction and the different types of training and testing samples. The experimental results showed that the back-propagation (BP) network structure of 3-14-6 and 6 principal components were the best choices. The consistency between training samples and scanned objects was important for the spectral reconstruction of images based on color scanner.