Intelligent thermography for detecting melamine adulteration in powdered milk

Abstract

A non-destructive, fast, inexpensive, and accurate technique was developed to detect and semiquantify melamine in powdered milk using a combination of infrared thermography and convolutional neural networks, specifically the ResNet34 architecture. Three types of powdered milk were mixed with varying melamine concentrations (0.5–10 ppm), each prepared in triplicate to ensure reproducibility. After heating, nearly 28,500 thermographic images were collected during the cooling process. Ninety percent of the dataset was used for training and internal validation of the convolutional neural networks, while the remaining 10 % was reserved for blind testing. The resulting algorithm classified thermographic images by milk type and melamine concentration with an overall accuracy exceeding 98 %. This performance highlights the potential of the proposed method as a reliable tool for detecting food fraud and ensuring dairy product quality, with an empirical detection limit of 0.5 ppm, well below the regulatory threshold for infant milk formulas.