Non‐Invasive Detection of Foreign Matter Embedded Inside Egg Fried Rice Using Hyperspectral Transmittance Imaging

ABSTRACT The mass production of egg fried rice (EFR) may unavoidably introduce foreign matter (FM), thereby jeopardizing consumer health and corporate reputation. However, typically embedded inside EFR, FM is challenging to be identified with computer vision and reflectance spectroscopy techniques. Additionally, internal FM detection methods including metal detectors and x‐ray systems struggle to detect FM with specific physical properties. In this study, hyperspectral transmittance imaging (HSTI) was employed to detect and localize FM sandwiched inside EFR based on their compositional differences. To enhance the FM classification and visualization outcomes, the imaging parameters of band range (VNIR: 371.57–1028.85 nm, SWIR: 899.24–1390.07 nm) and light intensity (70, 90, and 110 W) were thoroughly compared. Following the Savitzky–Golay (SG) smoothing plus standard normal variate (SNV) pretreatment, spectral data were downscaled by principal component analysis (PCA) and competitive adaptive reweighted sampling (CARS). Feeding the raw and screened spectral data into the three models containing K ‐nearest neighbor (KNN), random forest (RF), and multilayer perceptron (MLP), the results showed that VNIR‐based detection generally outperformed SWIR. Under 110 W light intensity in the VNIR band, the PCA‐downscaled spectra with MLP models on the test sets performed best, with the accuracy, precision, recall, and F 1‐scores of 98.33%, 98.59%, 98.53%, and 98.52%, respectively. Furthermore, the predictive visualization maps showed distinct FM locations, with their distributional accuracies aligning with the model outcomes. This study explored ideal protocols for FM detection embedded in EFR based on the HSTI systems, setting an example for accurate and comprehensive FM detection in the rice‐based food industry.