Detection of adulterants in pepper powder using deep learning-based microscopic imaging: A cost-effective approach for food safety

High-valued black pepper-based products are adulterated with papaya (Carica papaya) seeds, and chilli (Capsicum annuum) seeds. Those are causes of food allergic reactions and food poisonings. Therefore, a cost-effective and efficient system should be developed to identify those adulterants. According to studies, each plant material has its own unique microscopic image fingerprint. Phase contrast microscopes are able to differentiate those features based on their refraction indexes. This research was developed for identifying the best deep learning architecture to differentiate adulterated black pepper powder from genuine black pepper powder by using microscopic images. TensorFlow deep learning models (InceptionV3, Inception ResNetV2, ConvNeXtLarge, Xception, VGG-19, and ResNet-50) were used as the backbone. The Adam optimizer and categorical cross-entropy as a loss function were used in both the evaluation stages and the training stages. Categorical Accuracy, Precision, F1Score, False Negatives, False Positives, True Negatives, and True Positives were used for evaluating the trained models. 0.001 and 0.0001 Fine-tuning with learning rates were used for model training. According to the results of evaluation matrices, InceptionV3, Xception, and ResNet-50 showed superior performances in differentiating genuine black pepper powder from other powder classes. But overfitting was identified when increasing the number of training epochs at lower learning rates. Therefore image segmentation should be introduced for overcoming the challenges in whole image labelling. Selective hierarchical neural network layer fine-tuning and hyper parameter tuning can be implemented for overcoming the model training based issues. According to the findings, those microscopic image based techniques are able to identify genuine black pepper powder from other adulterated black pepper powder.