Researchers in China have developed a lightweight deep learning system for rapid, non-destructive analysis of wheat flour composition.
A recent study conducted by researchers from Henan University of Technology and Northwestern Polytechnical University looked to investigate a new method for food quality monitoring. This method involved using near-infrared (NIR) spectroscopy spectral analysis with artificial intelligence (AI). This study, published in Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, was led by Quan Pan of Henan University of Technology and proposed a lightweight convolutional neural network (CNN) model capable of real-time, non-destructive analysis of wheat flour quality using NIR spectroscopy (1).
Wheat flour has been a domestic crop for thousands of years (2). It is a powder that is made from grinding wheat, and it is used for various purposes. For one, it is an important ingredient in many food products, including bread, cookies, and cakes, to name a few (2). As a result, consumers are invested in the quality of wheat flour. Two factors normally assess the quality of wheat flour, and that is the protein and moisture content (2). Conventional methods for assessing these qualities are typically accurate, but they are often labor-intensive, time-consuming, and impractical for real-time or large-scale processing. As a result, these limitations have posed a significant challenge to industries striving to maintain quality while scaling up production (1).
In this study, Pan and the team designed a novel CNN model, called LGAKNet, that leverages several features. These additional features include ghost bottlenecks, external attention modules, and a Kolmogorov-Arnold network (1). These features were important to the study because they helped significantly improve the model's ability to extract relevant features from spectral data, enabling it to predict protein and moisture content in wheat flour more accurately (1).
The research team found that the model achieved an R² value of 0.9653 for protein content prediction, with a root mean square error (RMSE) of 0.2886 g/100 g and a residual predictive deviation (RPD) of 5.8981 (1). For moisture content, the R² was slightly higher at 0.9683, with an RMSE of 0.3061 g/100 g and RPD of 5.1046—strong indicators of the model's predictive power and reliability (1).
As a result, the researchers demonstrated the practicality of the LGAKNet model. Because the LGAKNet model can perform rapid and highly accurate flour analysis, it is ideal for integration into online monitoring systems used in food manufacturing environments (1). Its lightweight architecture ensures efficient operation without the need for high-end computing resources, making it suitable for deployment in edge devices and factory-floor applications (1).
To develop and validate their model, the researchers compiled an extensive data set from 519 flour samples derived from ten different wheat flour varieties (1). These samples were sourced from major flour producers across three major producers: China, the United States, and Russia (1). The varieties represented three different gluten content levels: high, medium, and low, which allowed researchers to test the model across a wide range of flour types.
Then, these 519 flour samples were analyzed using NIR spectroscopy. After NIR analysis, these samples were divided into smaller portions to measure protein and moisture content through standard laboratory procedures (1). The average values from these measurements were used as reference data for training and validating the CNN model (1). This data set enabled the development of a reliable regression model capable of predicting key compositional parameters from NIR data alone.
Summarizing their findings, the research team suggests that the LGAKNet model could be adapted for analyzing other food products and additional quality parameters (1). They also suggest that future studies should incorporate advanced spectral preprocessing techniques and transfer learning strategies to improve adaptability in diverse settings (1).
As a result, this study shows another example of how AI is being used in the food analysis industry. As AI continues to be increasingly used in the food analysis industry, we can expect that quality control processes will become faster and more accurate as the technology develops. The goal in food analysis remains reducing waste, enhancing consistency, and ultimately improving consumer trust in food products, and spectroscopy is helping to realize this vision.
