Pixel-level spectral reconstruction and compressed projection based on deep learning in detecting aflatoxin B1

Abstract

Aflatoxin, a highly toxic substance posing a substantial threat to food safety, necessitates a reliable detection method. This paper introduces a pioneering aflatoxin detection approach based on spectral reconstruction and projection compression model. The proposed method effectively addresses data imbalance by reconstructing aflatoxin spectra. The reconstructed spectra achieve remarkable performance, with a training RMSE (root-mean-square error) of 0.0242 and a test RMSE of 0.0214. Subsequently, the LSTM (Long Short Term Memory) model is trained on a dataset comprising 25% reconstructed AFB1 spectra and 75% original spectra, resulting in a testing accuracy of 98.55% and a testing loss of 0.0611. To further enhance the model performance, PCA (Principal Component Analysis) and compression projection are employed to reduce the LSTM model’s parameters. Despite reducing the LSTM internal parameters, the fine-tuned LSTM achieves an outstanding testing accuracy of 98.30%. This research presents a practical and efficient aflatoxin detection approach, offering improved accuracy and significantly reduced model complexity. The proposed algorithm holds great potential for enhancing the detection capabilities of intelligent sorting equipment.