(1) Selection of training samples: The number of training sam-ples is 1020, corresponding to the RGB values of 1020 color checkers.
(2) Data normalization: Since the orders of magnitude differbecause of the different variable units, the data shouldbe normalized before they are trained. In this paper the normalization function “mapminmax” is adopted. Data aremapped into [0.1, 0.9] to ensure convergence of the network.