Integrating Computer Vision and Deep Learning in Nutrition: A Novel Model for Multi-Label Ingredient Classification

Each year, poor diets contribute to more deaths in the United States than any other risk factor. Image classification has emerged as a promising opportunity to enhance food analysis capabilities for diet assessment and health monitoring. However, existing models are often limited to single-label classification due to a lack of ingredient-level data, hindering their applicability to food analysis tasks. In this work, we present a novel multi-label classification model powered by a ResNet-50 backbone. We trained a custom head on our self-curated dataset comprising 183 ingredient classes, using focal loss and threshold optimization to enhance classification performance. The model achieved 99.14% validation accuracy and reached a macro F1 score of 63.82% at an optimal threshold of 0.70. Our dataset and model provide a benchmark for further research in automated visual assessments of food items. This work can legitimize a new paradigm for AI-driven ingredient recognition as a foundation for data-driven dietary assessment.