The content of protein and fat in dairy products is determined by the near infrared transmittance spectroscopy （NITS）.The spectrum is preprocessed by methods such as smoothing，then，the data is compressed by wavelet function db3 and compression level 6.The prediction models of protein and fat in the four dairy products are established by radial basis function artificial neural network （RBF-ANN）using the compressed spectrum data as the inputs.The best spread value is obtained by experiments.When the spread is 135，the prediction accuracy of prediction set of protein is the highest and the Correlation coefficient and Mean square error are 0.9999 and 0.0301 respectively.In the same way，when the spread is 105，the prediction accuracy of prediction set of fat is the highest and the Correlation coefficient and Mean square error are 0.9997 and 0.0968 respectively.The results show that the model based on RBF-ANN combined with wavelet is more stable and more accurate.It can be used to test the qualities of dairy products quickly and nondestructively.