Study on Detection Methods for Quality Indicators of Japonica Rice Based on Hyperspectral Signals

Traditional chemical detection methods cause damage to samples and have low detection efficiency, while traditional machine learning algorithms cannot meet the needs of simultaneous multi-task learning. Based on Transformer and Convolutional Neural Network (CNN), this paper proposes a CNN-DecouplingTransformer model. The model is aimed at completing the simultaneous detection task of multiple quality information of japonica rice and improving the detection capability of the model. Experiments were conducted on a self-collected hyperspectral signal dataset of japonica rice. The experimental results show that the proposed model is more suitable for the detection of japonica rice quality information than traditional machine learning algorithms. Among them, the Rp²for moisture detection reaches 0.9331, which is significantly higher than the 0.9054 of the Transformer model. Meanwhile, compared with traditional machine learning algorithms, the proposed model has better simultaneous multi-index detection capability. When detecting multiple indicators simultaneously, the classification accuracy of imperfect grains still reaches $\mathbf{1. 0}$. This study provides a reliable and accurate method for the joint detection of multiple indicators of japonica rice quality.