Large-scale image classification and nutrient estimation for Chinese dishes

The convergence of nutritional science and machine learning has opened new avenues for improving dietary assessments and public health. This study addresses the challenge of generating accurate nutrition labels for diverse and complex Chinese dishes by leveraging advanced deep learning techniques. Utilizing the CNFOOD241 dataset, the by-far largest collection of Chinese dish images, we developed a robust classification and nutrient estimation model. Our approach integrates state-of-the-art models, including the RegNet Y series, and employs focal loss and data augmentation techniques such as cutmix and mixup. The model fusion achieved top 1 and top 5 classification accuracies of 86.28 % and 98.49 %, respectively, and demonstrated strong predictive performance for nutrient estimation with a mean R2-top5 value of 0.8636. This research offers a scalable and accurate method for generating nutrition labels for Chinese dishes, contributing significantly to the fields of food classification and nutritional analysis.