A YOLO-Based Framework for Automated Food Recognition and Nutrient Adequacy Estimation in the Free Nutritious Meal Program

Abstract

The Free Nutritious Meal (MBG) program in Indonesia may not consistently meet nutritional standards, providing an average of 377 of the targeted 531 calories per serving, with insufficient protein and fat. Limited nutritional information from food images makes independent meal assessment difficult. This study develops an image-based food detection system using You Only Look Once (YOLO) to estimate the nutritional adequacy of MBG meals based on Recommended Dietary Allowances (AKG). The study used Knowledge Discovery in Databases (KDD)- guided data management and Rapid Application Development (RAD)- guided system development. A dataset of 1.200 images across 51 food classes was evaluated using YOLO 8-YOLO 12, utilizing 5-fold cross-validation, followed by retraining and a 1.000 iteration bootstrap analysis on 240 test images. Average mAP50 scores were 0.8981, 0.898, 0.896, 0.894, and 0.888 for YOLO 8, YOLO 9, YOLO 11, YOLO 12, and YOLO 10. The Kruskal-Wallis test indicated significant differences among models (H = 489.946; p < 0.001), and YOLO 8 was selected because it had the highest bootstrap mean rank. Detected food classes were assigned reference serving weights rather than measured mass or volume, converted to nutrient values using the Indonesian Food Composition Table (TKPI), and compared with AKG by school level, age, and gender. Validation against nutritionist assessments across five nutrients produced a mean absolute percentage error of 40.31%, with energy and carbohydrate showing the closest agreement. A sensitivity analysis showed a 20 percent portion-weight error most strongly affected iron and calcium adequacy. Because nutrient values rely on reference assumptions rather than measured mass or volume, the outputs represent an estimate, not a definitive assessment of nutritional adequacy. YOLO 8 was the most suitable model and the highest mean rank assumption-based estimation approach for MBG nutritional screening, with direct mass or volume measurement recommended for future work.