Image-Based Detection of Adulterants in Milk Using Convolutional Neural Network

Adulteration of milk poses a severe human health hazard. Existing methods for detecting adulterants such as water, urea, ammonium sulfate (AmS), oils, and surfactants in milk are selective, expensive, and often challenging to implement in rural areas. The present work shows the potential of machine learning to detect milk adulterants using patterns of evaporative milk deposits. The final deposit patterns obtained after evaporation of the adulterated milk droplets are used to create an image data set. This data set is used to develop a deep learning model that deploys a convolutional neural network (CNN/ConvNet) to classify the distinct evaporation patterns obtained for different types and concentrations of adulterants. Further, we apply implicit and explicit regularization and compare their accuracies. The models trained with different regularization optimization schemes demonstrate that a CNN can be successfully implemented to detect adulterants in milk. Additionally, we experimentally determine how the type and concentration of milk adulterants, including ammonium sulfate (AmS), urea, oil, and surfactants, affect milk evaporative deposition. Added AmS and urea in milk crystallizes during evaporation to produce recognizable patterns that can be used for their detection. The method is capable of detecting AmS added in excess of 2.4% and urea in excess of 5% in diluted milk (20 wt %) due to the crystallization of AmS and urea, respectively. In the case of milk adulterated with vegetable oil, evaporation leads to the separation and accumulation of oil at the top of the deposit, leading to the detection of oil present in excess of 2% in 20% diluted milk. Furthermore, a minimum individual amount of 5% urea, 2.4% AmS, and 2% oil concentration in diluted milk (20%) is shown to be individually detected by evaporation pattern-based technique when milk is adulterated with all the adulterants (water, urea, AmS, and oil + surfactant) together. When subjected to different regularization optimization schemes, the CNN gives varying degrees of accuracy for successful detection. The use of implicit regularization in the form of data augmentation gives the best results with a testing average accuracy of 98%, showing that a CNN can be successfully deployed to classify and detect adulterants in milk.