Deep Learning-Based Hyperspectral Image Reconstruction from RGB Data for Gluten Detection and Quantification in Food Products

Detecting and quantifying gluten in food products is crucial for safeguarding individuals with gluten-related disorders, such as celiac disease and gluten intolerance. Effective detection technology will safeguard against harm to consumers (food safety) and reduce the potential losses of the industry. Hyperspectral imaging (HSI) has emerged as a powerful tool for gluten detection, owing to its ability to capture both spatial and spectral data. However, the high cost, complexity, and need for expert operation have limited its accessibility, particularly for consumer-level applications. This study addresses these challenges by investigating the deep learning techniques used to reconstruct hyperspectral images from standard RGB images, making gluten detection and quantification feasible for handheld devices. In this study, corn flour (CF) samples were contaminated with varying concentrations of wheat flour (WF), ranging from 0 to 10% at 0.1% (0-0.24%) and 0.5% (2.5-10%) increments. HSI data were captured with a camera covering the 400–1000 nm range, while RGB images were obtained using the built-in camera of a Samsung Galaxy smartphone. To identify the most relevant spectral regions for gluten detection, ground-truth hyperspectral cubes were constructed using key wavelengths selected by feature selection algorithms. The study compared different hyperspectral reconstruction algorithms, including Hyperspectral Convolutional Neural Network-Dense (HSCNN-D) and High-Resolution Network (HRNET) for gluten detection and quantification. Performance was evaluated using metrics such as the mean relative absolute error, root mean square error, and peak signal-to-noise ratio. K-nearest neighbors (KNN) and Random Forest (RF) classifiers were applied to detect gluten sources, and RF regression was used to quantify gluten sources. KNN and RF achieved testing accuracies of 85.7% and 90.8%, respectively, for the HRNET model and above 80% in the HSCNN-D and other models. RF regression showed an R²p of 0.6 for HRNET and values above 0.5 for the other models.