A dataset of 310 samples of tablet were obtained by using near infrared spectroscopy (NIR) technique, and then the NIR data were used to discriminate the four types of tablets with three scales. Wavelet clustering algorithm, a new unsupervised method, which applied a classical clustering strategy on the suitably chosen subset of wavelet coefficients, was introduced to improve the clustering performance. The optimal wavelet decomposition and wavelet coefficients partition were determined according to the index of discriminant accuracy. The total accuracy rates for laboratory-scale, pilot-scale and full-scale tablets samples were 100%, 100% and 99.2%, respectively, with only one sample misclassified. The overall results indicated that the wavelet clustering was an effective way for the discrimination analysis. NIR combined with wavelet clustering method is surely much more rapid and easier to use, and offers a feasible solution to the quality control of pharmaceutical tablet products.