To establish a quantitative detection model of benzoic acid additive in flour, terahertz time-domain spectra of benzoic acid doped at different percentages (mass fraction) in flour are collected, and the absorption coefficient spectra are obtained through calculation. It is found that the absorption peak amplitude is positively correlated with benzoic acid content. As for the detection method, first, we explore the effects of different spectral pretreatment methods on THz spectroscopy, and then adopt methods like smoothing correction, multiple scatter correction (MSC), baseline correction, and normalization correction to perform the appropriate processing. Subsequent to correction, PLS model is established to select the optimal pretreatment method. Experimental results verify that PLS model established subsequent to normalization is more optimal, with the correlation coefficient of prediction (r_p) observed to be 0.9790 and root-mean-square error of prediction (RMSEP) observed to be 1.28%. We establish PLS, least squares support vector machine (LS-SVM), and back propagation neural network (BPNN) regression models for the determination of benzoic acid content in flour. It is proved that the most optimal quantitative determination model of benzoic acid content in flour is BPNN model with correlation coefficient of prediction (r_p) of 0.9945 and root-mean-square error of prediction (RMSEP) of 0.66% subsequent to the normalization of terahertz absorption coefficient. It is concluded that a new solution for the nondestructive detection of benzoic acid additives in flour has been developed, and provide guidance for the detection of other types of additives, all of which are essential for the healthy development of the flour industry.