This study (open access) aimed to develop a non-invasive computer-vision method based on deep learning to detect water adulteration in orange juice. The model was trained on three juices (freshly squeezed juice, juice from concentrate, and orange nectar ) with water-adulteration levels ranging from 1% to 15%
The authors repeated the study with different shutter-speed-based exposure conditions. Training data were analyzed using ResNet50 convolutional neural networks. The best classification accuracy (88%) was found with images acquired at 1/250 s
The model was then evaluated using 240 independently prepared and blindly coded samples obtained from subsequent purchases of the same three commercial products. This independent blind validation achieved a 24-class accuracy of 86.7%. Misclassifications occurred predominantly between adjacent or closely related water-adulteration levels within the same juice product. The authors report that when the independent-validation predictions were collapsed into a binary pure-versus-adulterated screening task, the model achieved 100.0% sensitivity, 93.3% specificity, 99.2% accuracy, and 96.7% balanced accuracy.
They conclude that the method is feasible as a rapid pre-screening tool for detecting visible-image patterns associated with controlled water dilution when there is no need to discriminate between different levels of adulteration. The independent validation provides evidence of transferability to newly prepared samples from subsequent purchases of the same products. However, further validation across additional brands, production batches, orange origins, seasons, and acquisition environments is required before broader applicability can be established.
Photo by Dmitry Ganin on Unsplash
Comments