In order to use the near infrared spectroscopy to rapidly detect the concentrations of fat, protein and lactose in the milk, a preprocessing method of weight resetting for near infrared spectroscopy based on histogram normalization and fuzzy analytic hierarchy process is proposed. The principal component analysis is conducted for the spectrum of milk samples, the best principal component number in the spectral data is determined, and the scores and weights of the principal components are obtained. Filtering and purification and noise elimination of two-dimensional spectral matrix are realized by using the mathematical statistic idea of histogram normalization. The fuzzy analytic hierarchy process is used to reset the weight of the active principal component information, and the irrelevant interference information of principal component is filtered out, thereby, the spectrum is reconstructed. On this basis, partial least squares 1 (PLS1) regression models of fats, proteins and lactose are built after spectral data preprocessing, getting the correlation coefficient of fat is 0.980 and predicted root mean square error is 0.158×10-2 g·mL-1; the correlation coefficient of protein is 0.997 and predicted root mean square error is 0.050×10-2 g·mL-1; the correlation coefficient of lactose is 0.985 and predicted root mean square error is 0.152×10-2 g·mL-1. Through the model prediction results, we can see that this pretreatment method has better filtering and noise elimination effect than the conventional pretreatment method. It is feasible to pretreat the near infrared spectrum of milk by combining histogram normalization and fuzzy analytic hierarchy process.