Detection of food adulteration using machine learning tools

Abstract

Adulteration of food has been an urgent concern for the health and food safety of the people, and the affected area is particularly the developing countries due to the inefficient infrastructure and limited resources available to test the food items. Some traditional forms of detection, like chemical analysis and chromatographic analysis, are usually slow, costly, and require competent manpower. This paper explores an affordable and scalable Machine Learning (ML) system to identify food adulteration in sugar and poppy seed, and also another criteria is to identify the freshness of meat. The methodology of the study uses a modular approach to using self created RGB image dataset of sugar mixed with Magnesium Sulfate as adulterant which is one of its first kind in food adulteration detection of sugar and another one is of poppy seed dataset which is mixed with semolina as adulterant, collected through online to predict different fixed adulteration levels. Additionally, the dataset of meat is collected from an existing paper for further working on meat freshness detection. A self proposed ensembled model using pretrained models including: Densenet121, Conv-Next Tiny and Resnet50 with soft voting is used for detecting adulterants of sugar with Grad-CAM visualization, also showed some success in detecting other unknown adulterants. The ensembled model for sugar has shown about 88.96% accuracy. Whereas for poppyseed, Convolutional Neural Networks (CNN), and its pretrained model- EfficientNetB0 shows very high accuracy - 98.25% itself to measure the level of adulterants in class-wise manner without requiring any ensemble learning. As per previous works, there are few models that have satisfactory accuracy rate in terms of adulterant detection but there still exist issues, especially in covering diverse food items like sugar and poppyseed and as of for meat freshness, the Misclassification Cost based on some shown pretrained models was much higher in a previous work, which has been addressed by this study. For meat dataset, a Swin Transformer integrated with Explainable AI succesfully identified spots of spoilage, solving the black-box problem, showed increased accuracy upto 98% which is higher compared to that of the models shown in the existing paper from where the meat data set was collected and along with this, the study addressed the misclassification cost, an evalutation metric proposed by the original paper, which was lowered by many folds compared to the exisiting misclassification cost of the original paper. This research highlights that image data by mobile or obtained on accessible platform with AI can facilitate fast, dependable food safety testing and aid food quality inspection in supply chains.