Uso de infravermelho próximo e visão computacional combinados com aprendizado de máquina para classificação de fubá orgânico adulterado

Abstract

The growing demand for organic products and their high added value have driven an increase in food adulteration cases, compromising product authenticity and consumer trust. This study proposed the development and evaluation of two non-destructive methods for detecting fraud in organic cornmeal adulterated with wheat flour: computer vision through analysis of color channels (RGB/HSV) and near-infrared (NIR) spectroscopy, both combined with a machine learning algorithm, Random Forest (RF). Samples of cornmeal with two levels of adulteration (20% and 30%) were prepared, along with a set of pure, non-adulterated cornmeal samples. The samples were photographed, and both color attributes and NIR spectral readings were extracted. The classification models were developed using Python and R, with validation to assess their classification performance. The evaluation metrics were presented using confusion matrices and F1-scores. The color-based model achieved an accuracy of up to 75% on the test set, with the saturation and red channels standing out as the most relevant. The spectroscopic model achieved an accuracy greater than 99%, demonstrating high precision and excellent generalization capability. The results confirm the effectiveness of the proposed techniques, which proved to be viable, accessible, and promising for the authentication of food products and can be applied in routine quality control in the food industry or regulatory inspections.The growing demand for organic products and their high added value have driven cases of food adulteration, compromising authenticity and consumer confidence. This work proposed the development and evaluation of two non-destructive methods for detecting fraud in organic cornmeal adulterated with wheat flour: computer vision by color channel analysis (RGB/HSV) and near-infrared spectroscopy (NIR), both combined with machine learning algorithms. Cornmeal samples with three levels of adulteration (0%, 20% and 30%) were prepared and subjected to the extraction of chromatic and spectral attributes. The classification models were built in Python and R, with validation by accuracy, confusion matrix and F1-score. The model based on color attributes showed accuracy of up to 75% in the test set, with emphasis on the saturation and red channels as the most relevant. The spectroscopic model achieved accuracy above 99%, with high precision and demonstrating excellent generalization capacity. The results prove the effectiveness of the proposed techniques, which proved to be viable, accessible and promising for the authentication of food products, and can be applied in quality control routines in the industry or in regulatory