Residual neural networks to quantify traces of melamine in yogurts through image deconvolution

Abstract

This work combines intelligent algorithms based on transfer learning and residual neural networks (viz., ResNet34) to process and classify optical images of pure and adulterated yogurt samples. This integration aims to detect the presence of melamine in yogurt in concentrations ranging from 1 to 10 ppm. An image database of 1,888 images is used to train the ResNet34, and 212 blinded images to test and validate its performance. The optimized intelligent algorithm is able to classify the images into 21 classes considering the yogurt type and melamine content, obtaining an accuracy of over 94%. These encouraging results certify a simple yet powerful real-time quality control method for producers and distributers to ensure food safety for the final consumers, while pinpointing the source of potential fraudulent procedures.