Deep Neural Network-Based Sorghum Adulteration Detection in Baijiu Brewing

Abstract

In Baijiu brewing process, suppliers may adulterate glutinous sorghum with commercially inferior japonica sorghum, which can affect the yield and quality of the final Baijiu production. Currently, sorghum adulteration detection in Baijiu brewing process in China is carried out manually by sampling and observation, which strongly depends on the experiences of workers. In this paper, we proposed a method that uses sorghum images as input and combines image processing and deep neural networks to identify grain varieties and calculate the adulteration ratio. Two derivative networks of CNN, i.e., ResNet and SqueezeNet are used to implement the deep neural networks for sorghum grain identification and adulteration ratio calculation. The classification accuracy of the ResNet and SqueezeNet based models reached 93.34% and 87.98% on test set, respectively. The root mean squared error (RSME) for adulteration ratio estimation is 4.95% and 7.73%, respectively. The mean absolute error (MAE) is 4.20% and 6.29% accordingly. The proposed pipeline is capable of realizing rapid and non-destructive adulteration detection of raw materials in industrial production, thus conducing to the industrial digital transformation and efficiency improvement.

Keywords

• Adulteration detection
• brewing process
• deep learning
• image processing
• machine vision