Peanut is an important oil crop that is susceptible to Aspergillus infection to produce aflatoxins, of which Aflatoxin B₁ (AFB₁) poses a higher threat to humans and animals. The traditional AFB₁ detection method is more complicated, such as cumbersome operation, the material is destroyed and long time-consuming. Therefore, it is of great significance to develop a rapid, non-destructive and suitable online detection method for peanut production and processing. Firstly, peanuts were purchased and stored at 28 ℃ and 85% relative humidity to mold. In the 0, 4, 6, 7 and 8 d time periods, the spectral and image information of peanut samples were collected dynamically at a rate of 0.15 m·s^-1. After collecting the information, the AFB₁ content in peanuts was determined by enzyme-linked immunosorbent assay (ELISA). Data processing steps are: pre-processing of spectra by multiple scatter correction, baseline correction, standard normal variable correction, and Savitzky-Golay smoothing, principal component analysis of spectral data in the range of 600~1 600 nm, 8 characteristic wavelengths (630, 1 067, 1 150, 1 227, 1 390 and 1 415 nm) are determined according to the principal component weight coefficient; the image is subjected to grayscale and threshold segmentation processing, and 12 image color feature parameters are extracted. Finally, using linear discriminant analysis (LDA) and support vector machine (SVM) to establish a qualitative discriminant analysis model for peanut samples (with a national standard of 20 μg·kg^-1). ELISA results showed that AFB₁ peanut exceeding 58%. Visible/near-infrared spectrum analysis showed that the absorbance gradually decreased with the increase of toxin intensification at the peak of 1 180 nm. Machine vision analysis showed that with the increase of storage time, the RGB value of peanuts decreased overall, the surface gradually dimmed and covered with hyphae, and the level of toxin infection gradually increased. The principal component analysis showed that the spectrum showed a clear clustering trend, but the clustering trend of image and data fusion was not obvious. LDA and SVM models constructed based on full-spectrum and characteristic wavelengths can quickly identify over- and under-standard samples, the best recognition rate based on the full-spectrum model is 92%, and the best recognition rate based on the characteristic wavelength is 88%; compared with spectral information modeling, the nonlinear SVM model performs better on image color feature parameter modeling , and the best recognition rate is 90%; combining the internal and external information of peanut samples, the SVM model based on spectral and image information fusion has the best discrimination accuracy of 92%. It is feasible to use the visible/near-infrared spectroscopy and machine vision technology combined with stoichiometry to realize the dynamic discrimination of peanut AFB₁ content exceeding the standard, which provides a theoretical basis for the quality and safety detection of peanut.