Establishment of a Daqu Grade Classification Model Based on Computer Vision and Machine Learning

Mengke Zhao, Chaoyue Han, Tinghui Xue, Chao Ren, Xiao Nie, Xu Jing, Haiyong Hao, Qifang Liu, Liyan Jia

The grade of Daqu significantly influences the quality of Baijiu. To address the issues of high subjectivity, substantial labor costs, and low detection efficiency in Daqu grade evaluation, this study focused on light-flavor Daqu and proposed a two-layer classification structure model based on computer vision and machine learning. Target images were extracted using three image segmentation methods: threshold segmentation, morphological fusion, and K-means clustering. Feature factors were selected through methods including mean decrease accuracy based on random forest (RF-MDA), recursive feature elimination (RFE), LASSO regression, and ridge regression. The Daqu grade evaluation model was constructed using support vector machine (SVM), logistic regression (LR), random forest (RF), k-nearest neighbor (KNN), and a stacking model. The results indicated the following: (1) In terms of image segmentation performance, the morphological fusion method achieved an accuracy, precision, recall, F1-score, and AUC of 96.67%, 95.00%, 95.00%, 0.95, and 0.96, respectively. (2) For the classification of Daqu-P, Daqu-F, and Daqu-S, RF models performed best, achieving an accuracy, precision, recall, F1-score, and AUC of 96.67%, 97.50%, 97.50%, 0.97, and 0.99, respectively. (3) In distinguishing Daqu-P from Daqu-F, the combination of the RF-MDA method and the stacking model demonstrated the best performance, with an accuracy, precision, recall, F1-score, and AUC of 90.00%, 94.44%, 85.00%, 0.89, and 0.95, respectively. This study provides theoretical and technical support for efficient and objective Daqu grade evaluation.