In the present paper, a kind of noninvasive determination for human serum protein concentration of albumin, globulin and total protein was explored based on the technology of near-infrared reflectance spectra. Reflectance spectra on the tongue tip of 58 volunteers were collected. Because these is an nonlinear mapping relationship induced by the individual differences between these spectra data and serum protein concentration, SVM was used to establish quantitative regression models of 3 kinds of protein concentration respectively after the normalized spectral reflectance was calculated and the protein content statistics distribution of the sample set was analyzed. In addition, results of SVM were compared with that of PLS. The results show that the predictive effect for calibrated model of SVM is obviously better than that of PLS. Using SVM model to predict the prediction set, the correlation coefficients of ALB, GLB and TP are respectively 0.894, 0.931 and 0.863, and root mean square errors are 2.19, 1.93 and 4.38. So SVM can resist the impact of nonlinear factors among in-vivo determinations, and enhance the robustness of the models. Meanwhile it was also showed that the near infrared spectral information for tongue can relatively objectively reflect changes in human physiological and biochemical indexes, and this technology for noninvasive determination of serum protein concentration is highly feasible.<作者简国家自然科学基金项目（30973964）, 天津市应用基础及前沿技术研究计划项目（11JCZDJC17100）, 天津市科技计划项目: 科技型中小企业创新基金项目（10ZXCXSY10400）资助介>李家星, 女, 1979年生, 天津大学精密仪器与光电子工程学院在读博士