(P13-022-25) Machine Learning and Generative AI in Food Image Analysis: A Systematic Review on Applications for Reducing Food Waste and Enhancing Dietary Quality

Objectives: To assess the role of machine learning (ML) and generative artificial intelligence (AI) in food recognition and portion size estimation, focusing on their applications in reducing food waste and improving dietary quality. Specific aims include evaluating the performance, scalability, and integration challenges of these technologies in food services.

Methods: A systematic review following PRISMA guidelines analyzed studies from PubMed, Scopus, and Embase on ML (e.g., SVMs) and AI models (e.g., CNNs, Transformers, GANs) for food segmentation, classification, recognition, and volume estimation, extracting data on methods, performance metrics, and real-world applications.

Results: Among 82 studies, food segmentation was performed in 68%, food classification in 81%, one-shot food recognition in 21%, and volume estimation in 55%, with 43% utilizing 3D modeling techniques. Calorie and macronutrient estimation featured in 62% of cases. CNNs dominated (99%), while transformers and multi-task learning were used in 16%, GANs and Transfer Learning in 4%, and SVMs in under 2%. Challenges included reliance on GPUs (37%), limiting scalability, and sensitivity to real-world conditions (77%). Region-specific databases, like MedGRFood for Mediterranean and UECFOOD-256 for Asian foods, appeared in 34% of studies.

Food segmentation reached 74% accuracy, while classification achieved top-1 accuracy of 99.16%. Volume estimation models had relative errors between 3.73% and 17.45%. Multi-task models combining classification and volume estimation achieved up to 88.67% accuracy, with relative errors of 25.5%. CNNs dominated, exceeding 95% accuracy on datasets like Food-101, but challenges remain, including sensitivity to lighting and limited generalizability.

Conclusions: Food image recognition relies on CNN-based architectures and 3D modeling but faces challenges like lighting sensitivity, irregular food shapes, and limited dataset generalizability. Future research should expand datasets, integrate multi-modal inputs, and develop robust architectures to improve scalability and real-world applicability, addressing diverse global dietary needs.

Funding Sources: No funding sources were reported for this study.