In order to find a rapid detection method for the content of two main allergens (α_s1 and κ-casein) in milk, 211 Chinese Holstein milk samples from four provinces of Henan, Hubei, Ningxia and Inner Mongolia were selected as the research objects, and a non-destructive and rapid detection model of α_s1 and κ-casein in milk was established based on Fourier transform mid infrared spectroscopy. Firstly, the original spectrum of milk was pre-analyzed, and it was found that strongly influenced the spectral absorption of milk. The two main absorption regions of water (1 597~1 712 and 3 024~3 680 cm^-1) were analyzed. It was found that the absorption region of water (1 597~1 712 cm^-1) overlapped with that of protein (1 558~1 705 cm^-1)(amide Ⅰ). By comparing the effect of removing 1 597~1 712 cm^-1, the spectral region of 925.92~3 005.382 cm^-1 was selected as the sensitive band for subsequent analysis. The dimension of the selected full spectrum was reduced manually, and MCCV eliminated the abnormal samples. The support vector machine regression model (SVR) was established by using eight preprocessing algorithms, such as Savitzky-Golay convolution smoothing (S-G), standard normal variable (SNV). Meanwhile, three feature selection algorithms were combined, such as competitive adaptive reweighting algorithm (CARS) and information-free variable elimination algorithm (UVE). The results showed that for α_s1- casein, the SVR model established by the combination of the first derivative and CARS algorithm was the best, the training set correlation coefficient (R_C) and test set correlation coefficient (R_P) were 0.882 7 and 0.899 8, respectively, and the training set root mean square error (RMSEC) and test set root mean square error (RMSEP) were 1.136 3 and 1.372 6, respectively. For κ-casein, the SVR model established by the combination of second-order difference and UVE algorithm was the best. The training set correlation coefficient (R_C) and test set correlation coefficient (R_P) were 0.914 7 and 0.887 7, respectively, and the training set root mean square error (RMSEC) and test set root mean square error (RMSEP) were 0.473 5 and 0.558 1, respectively. The results showed that the SVR model based on Fourier transform mid-infrared spectroscopy can be used to detect the content of allergens α_s1 and κ-casein in milk, and the prediction effect was good. This study can make up for the blank of rapid and non-destructive detection of casein in milk by spectral technology in China.