In order to solve the problem of the influence of noise on model accuracy and stability in detecting materials content using near infrared spectra, we introduce the generalized S transform and singular value decomposition (SVD). Firstly, we use the generalized S transform to obtain time-frequency spectra of spectral data, and then use the two-dimensional time-frequency coefficient matrix as the Hankel matrix of SVD to solve singular values. Secondly, we use the k-means clustering algorithm to classify the singular value sequence and determine the reconstructed singular values. Finally, the de-noised coefficient matrix is transformed by the generalized S inversion to obtain de-noised spectral data. The basic theory and realization process of the combination method are given, and simulated data and the first derivative spectrum of wheat gluten are de-noised with the combination method. The results are compared with the traditional 9-point smoothing method and wavelet soft thresholding method. It is found that the proposed method overcomes the limitation of single dimension filtering (time domain or frequency domain), and does not need to reference noise data and select the base function. In the de-noising of wheat gluten derivative spectra, only 2 singular values are enough to achieve better de-noising effect, which reduces the complexity of the filtering process. The accuracy of the near infrared spectrum analysis and the robustness of the proposed model are better than those of the traditional 9-point smoothing method and the wavelet soft thresholding method. The predictive coefficient of the prediction set of the proposed method is 0.9985, which is larger than that of the 9-point smoothing method (0.9436). The root mean square error of the proposed method is 0.0406, which is smaller than that of the 9-point smoothing method (0.0843). The accuracy of quantitative detection of moisture content in wheat gluten is improved obviously.