In the field of noninvasive blood glucose sensing by near-infrared (NIR) spectroscopy, spectra are highly susceptible to the influence of background variations caused by the measuring instruments and physiological variations from the measured object because the concentration range of glucose in blood is usually small. It should be noted that the influence of background variations cannot be entirely removed; reasonable methods should be adopted in order to reduce the consequence of background variations to an acceptable level. One of the most common methods is to select a relatively stable standard material which shows similar optical property to the measured objects as the reference to perform the measurement. In order to maximize the elimination effect on the influence of background variations and realize a relatively accurate extraction of the glucose concentration information, a reference measurement method combined with double-beam spectra collection and net analyte signal (NAS) processing is proposed in this article. Spectra of the measured samples and reference substances are collected simultaneously with a double-beam double-detector measuring system. The NASs of glucose are obtained by projecting every spectrum of measured samples on the noise background subspace spanned by the spectrum of the reference substance. Experiments are conducted in pure absorption and strong scattering medium respectively. Two-dimensional correlation spectroscopy (2DCOS) and partial least squares regression (PLSR) are adopted as data processing methods to test the effectiveness of the reference measurement method. Results of 2DCOS show that the specificity of glucose concentration information in samples can be improved to a large extent by NAS processing compared to the reference subtraction method. Meanwhile, the root-mean-square error of prediction (RMSEP) of PLS model predicting glucose concentration decreases by 35.25% and 37.95% respectively in the double-beam experiment of glucose aqueous solution and 20%-intralipid solution compared with those in the single-beam experiments. These two RMSEPs further decrease by 26.11% and 14.84% after combining NAS processing with the double-beam experiment data. These results prove that the method of combining double-beam measurement with NAS processing is effective in extracting the glucose information and improving the accuracy of the calibration model, which provides more possibilities for noninvasive blood glucose sensing.