首页|Single-digit ppm quantification of melamine in powdered milk driven by computer vision
Single-digit ppm quantification of melamine in powdered milk driven by computer vision
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
Elsevier
In this work, a method to detect and quantify melamine in three different powdered milks is presented. The goal has been achieved by training convolutional neural networks (CNN) with a photograph database. The three types of milk are of different brands, each with their own fat content and intended final consumer (age-based distinction). The adulterated samples were prepared by weighing and adding trace amounts of melamine, even reaching samples that are considered to be "melamine-free". A total of 3100 images were taken to develop the CNNs (100 images per group to be classified). Specifically, a ResNet34 model architecture has been used to carry out the classification. For this deep learning approach, the images were randomly divided into two main sets: 90% for the training-validation phase of the CNN and 10% to serve as a blind test. The optimized model showed an overall accuracy of 98.7% during the validation phase, while leading to a 3.0% misclassification rate during blind testing, denoting the effectiveness of the application as a quality and safety control method for the milk industry.
NSTL
Elsevier
