A portable NIR-system for mixture powdery food analysis using deep learning

The combination of near-infrared spectroscopy and machine intelligence has been an emerging nondestructive tool for powdery food evaluation. In this research, a novel portable system (defined as NIR-Spoon) was presented for simultaneously evaluating the mixing proportion of multi-mixture powdery food. Convolutional neural networks for multi-regression (CNN-MR) and that for feature selection (CNN-FS) were proposed for spectra processing. Multi-mixture powder samples, which contained one or more components including milk, rice, corn and wheat, were inspected by the NIR-Spoon. Results showed that the partial least squares regression (PLSR) model estimated the proportion of mixture with root mean square error (RMSE) of 0.059 and correlation coefficient (R2) of 0.938. The proposed CNN-MR realized a further improvement comparing to the benchmark PLSR method, with 0.035 for RMSE and 0.976 for R2. The CNN-MR still kept R2 of 0.970 based on 25 features selected by the CNN-FS algorithm. Moreover, the integrated load sensor could convert the proportion to the weight of each component. All hardware and software were integrated on the NIR-Spoon. Overall, the NIR-Spoon provided satisfactory accuracy and user-friendly mobile applications. It also has excellent potential to be extended for inspecting other kinds of food products in future research.

Automated summary

The combination of near-infrared spectroscopy and machine intelligence is a promising tool for evaluating powdery food, and the novel NIR-Spoon system demonstrated good accuracy and portability in assessing the mixing proportion of multi-mixture powdery food. The use of convolutional neural networks for spectra processing can improve model accuracy, showing potential for future research in inspecting various types of food products.