Starch content is an important trait of maize (Zea mays L.) kernels as it accounts for the seed yield and quality. Analysis starch content accurately at the population level is the important foundation when we study genetic and physiological of starch quality. In this paper, 230 maize inbred lines were set as samples, using the method of polarimeter and pre-treatment of the first derivative add minus one line separately to establish and optimize a Near-infrared spectroscopy (NIRS) model of maize starch content successfully, which can improve the accuracy of the prediction significantly. Of the model, the calibration standard deviation (RMSEE) is 0.609, the cross-validation standard deviation (RMSECV) is 0.722, the external verification standard deviation (RMSEP) is 0.738, the calibration correlation coefficient (R2cal) is 0.909, the cross-validation correlation coefficient (R2cv) is 0.864, and the external verification correlation coefficient (R2cv) is 0.854. Of the model, the deviation between the predicted value and the chemical value can be controlled within 1.7%, which can improve the accuracy largely when it was used in quantitative analysis of grain starch content and then can be applicated in breeding inbred line selection or crude starch content analysis at the group level.