A multi-scale attention multi-task 1D-CNN for predicting rice chemical components at different milling degrees using hyperspectral imaging

Precise quantification of nutritional components in rice is crucial for achieving moderate milling. However, traditional detection methods often struggle to simultaneously meet the demands for efficient, non-destructive and multi-indicator analysis. To predict the chemical components of rice at different milling degrees, this study utilized a fusion of hyperspectral imaging and deep learning. The results demonstrated that the multi-task 1D-CNN with a multi-scale attention mechanism outperformed traditional regression models. The coefficient of determination and residual predictive deviation for cellulose, fat and protein exceeded 0.92 and 3.65, respectively. Additionally, the root mean square error and mean absolute error were below 0.16 and 0.13, respectively. The results introduce an innovative approach to food inspection and provide insights into the chemical variations associated with grain refinement.