The reasonability using near-infrared spectra (NIRS) for detection of melamine in liquid milk were studied. NIRS of samples were obtained by Fourier Transform Infrared Spectroscopy(FTIR) spectrometer, samples of liquid milk with adding melamine were analyzed while characteristic signals were extracted by wavelet transform and principal component analysis. Quantitative analysis was taken by General the Regression Neural Network(GRNN) for the diffuse reflectance spectroscopy of liquid milk adding melamine first. The result is not satisfactory enough that when concentration of melamine is lower than 0.05%, the R² of predict model is only 0.567. Qualitative discrimination by support vector machine was carried out later. The results showed that the model has a very good discriminate ability of the spectrum. The accuracy of the test set was 94.44%. Overall, NIRS combined with appropriate chemometrics methods is a feasible way to take qualitative discrimination of liquid milk whether or not adding melamine.