Pixel-level deep spectral features and unsupervised learning for detecting aflatoxin B1 on peanut kernels

Aflatoxin, with higher toxicity, is widely found in grains such as peanut and corn. This study proposes an unsupervised learning method to detect aflatoxin based on deep spectral features. The spectral feature clustering has the best effect based on Fuzzy C-means (FCM). This method clustering accuracy is 95.51%, and the verification accuracy is 98.36%. Then, the pre-trained model is applied to the limited spectral dataset, and the pre-trained network model (1-dimensional convolutional neural network) is trained from FCM clustering results. The random forest plus pre-trained model has the best performance, and the classification accuracy is 97.11%. Finally, we propose an aflatoxin content quantitative analysis method based on the clustering results. This study provides a new aflatoxin detection method, and it will facilitate advanced intelligent detection equipment development.