In order to quickly analyze varieties of orange juice with near infrared spectra, firstly, principal component analysis (PCA) was used to analysze the clustering of orange juice samples, and the characteristic differentia of four orange juice varieties was obtained through qualitative analysis. Then plentiful spectral data were compressed by wavelet transform (WT) and the model was built with radial basis function neural network (RBF-NN), which offered a quantitative analysis of orange juice varieties discrimination. The model regarded the compressed data as the input of RBF-NN input vectors and built a RBF-NN model. Two hundred forty samples from four varieties were selected randomly to build the training model, which in turn was used to predict the varieties of 60 unknown samples. The discrimination rate of 100% was achieved by WT-RBFNN method. It was indicated that wavelet transform combined with RBF-NN is an available method for variety discrimination based on the near infrared reflectance spectroscopy technology. It offered a new approach to the fast discrimination of varieties of orange juice.