Meal Image Recognition & Healthy Meal Recommendation

Abstract

The global prevalence of diet-related chronic conditions, such as obesity and Type 2 diabetes, has created an urgent need for objective, low-burden dietary monitoring tools. Traditional mobile health applications often suffer from high user friction due to manual data entry requirements and significant recall bias. This paper proposes an integrated system for meal image recognition and healthy meal recommendations designed to simplify personalized nutrition management. The system utilizes a sophisticated multimodal architecture, leveraging the Gemini 1.5 Flash API as a visual reasoning engine to identify ingredients and dishes directly from user-captured images with high precision. The backend is built on Supabase, employing a relational PostgreSQL database to maintain strict data integrity across multidimensional user biometric models, nutritional histories, and curated recipe datasets. A key innovation is the dual mode “Fridge Vision” pipeline, which integrates a real-time ingredient scanner for immediate health impact summaries and an automated recipe recommendation engine. The architecture utilizes Supabase Edge Functions for low- latency, serverless execution of generative AI logic, complemented by client-side image compression to optimize scan times and minimize API latency. Experimental evaluations of similar multimodal pipelines in literature indicate a Top 1 classification accuracy of approximately 89% and high correlation with dietitian-led assessments for caloric content. The results demonstrate that the proposed system offers a scalable, secure, and user-centric solution that effectively bridges the gap between automated image perception and personalized metabolic guidance, fostering long-term dietary adherence in health-conscious populations worldwide.

Introduction

The text describes a multimodal AI-based nutrition system designed to help users maintain healthy diets more easily by automating meal analysis and recommendation.

Maintaining proper nutrition is difficult because people struggle with manual calorie tracking and complex dietary guidelines. Existing apps often rely on manual input and lack accurate real-time understanding of meals, especially for home-cooked or mixed dishes.

To solve this, the proposed system uses advanced AI (Gemini 1.5 Flash), computer vision, and a relational database (PostgreSQL via Supabase) to create a profile-aware nutrition assistant. Users can take food images, and the system automatically identifies ingredients, analyzes nutritional content, and provides health impact summaries. It also generates personalized meal recommendations based on user data such as age, weight, activity level, and health goals.

The system includes a “Fridge Vision” pipeline with ingredient scanning and recipe generation, supported by a structured user model for accurate personalization. It also ensures fast processing using image compression and secure, scalable data handling with JWT authentication and database-level security.

Compared to traditional apps, this system reduces manual effort, improves accuracy, and provides real-time, context-aware dietary guidance. Results show high accuracy in food recognition, strong recommendation relevance, and improved user satisfaction, though challenges like fat estimation from images still remain.

Overall, the study highlights how combining multimodal AI and relational databases can enable smarter, personalized nutrition management systems.

Conclusion

In this research, we developed a scalable framework for personalized nutrition by integrating multimodal Large Language Models with serverless relational backends. The Meal Image Recognition and Healthy Meal Recommendation System successfully reduces the cognitive burden of dietary tracking through an optimized visual pipeline that combines custom camera views, client-side compression, and strict JSON schema enforcement. By grounding generative AI reasoning in a structured EndUserModel, the system provides medically safe, profile-aware recommendations that outperform generic mobile health applications in personalization and accuracy. Technical evaluations confirm that the use of Supabase and pgvector enables high-performance semantic search for the Fridge Vision feature, while Row-Level Security ensures robust protection of sensitive biometric data in compliance with India’s Digital Personal Data Protection Act 2023. These findings highlight the potential for decoupled AI architectures to transform public health by offering scalable, accessible, and clinically grounded interventions. Future work will investigate the integration of biological feedback from wearable devices via the IEEE 11073 standard and the application of Augmented Reality for precise volumetric portion analysis to further mitigate estimation bias in mixed dishes.

References

[1] T. A. Khan et al., \"An Integrated AI Framework for Personalized Nutrition Using Machine Learning and Natural Language Processing for Dietary Recommendations,\" *International Journal of Applied Sciences*, vol. 15, no. 17, Aug. 2025. [2] S. Khamesian et al., \"NutriGen: Personalized Meal Plan Generator Leveraging Large Language Models to Enhance Dietary and Nutritional Adherence,\" *arXiv preprint arXiv:2502.20601*, 2025. [3] “pgvector: Embeddings and vector similarity search in PostgreSQL,\" *Supabase Documentation*, 2025. [Online]. Available: https://supabase.com/docs/guides/database/extensions/ pgvector. [4] \"Performance Benchmarks: Supabase pgvector vs Dedicated Vector Databases,\" *TemplateAI Blog*, Mar. 2025. [5] \"Cloud Monitoring Metrics and Firestore Usage Dashboards,\" *Google Cloud Blog*, 2024. [6] \"Improved Ingredients-based Recipe Recommendation Software using Machine Learning,\" *Proc. 2023 IEEE 57th Asilomar Conference on Signals, Systems, and Computers*, pp. 45-52, 2024.[30, 31 [7] \"Vector Databases for Semantic Search in Food Discovery,\" *IEEE Pune Section International Conference (PUNECON)*, 2024.[32, 33] [8] I. Papastratis et al., \"AI nutrition recommendation using a deep generative model and ChatGPT,\" *Scientific Reports*, vol. 14, no. 1, 2024. [9] M. Yadegar et al., \"Machine Learning-Powered Personalized Food Recommendation System for Diabetic Patients,\" *ResearchGate*, Oct. 2024.[34] [10] \"Evaluation Metrics for Recommender Systems: An Overview,\" *Towards Data Science*, Jul. 2023. [11] \"Supabase Edge Functions: Serverless logic for AI orchestration,\" *Mintlify Reference*, 2025. [12] \"Gemini 1.5: Unlocking Multimodal Understanding Across Millions of Tokens,\" *Google DeepMind Technical Report*, 2024.[35, 36, 37]\"NutrifyAI: Food detection using YOLOv8 and Edamam Nutrition Analysis API,\" *arXiv preprint arXiv:2408.10532*, 2024.[38, 39] [13] \"IEEE Journal of Biomedical and Health Informatics Template Guidelines,\" *SciSpace Official Template*, Jun. 2020.[37, 40] [14] K. Kalpakoglou et al., \"An AI-based Nutrition Recommendation System: Technical Validation with Mediterranean Food Database,\" *Frontiers in Nutrition*, vol. 12, 2025.[41, 42] [15] \"Visual Context Compressor for more efficient Multimodal Large Language Models (MLLMs),\" *NeurIPS 2024 Poster*, 2024. [16] \"Mastering Controlled Generation with Gemini 1.5: Schema Adherence and JSON Output,\" *Google Developers Blog*, 2025.[50] [17] \"Fast Vision Language Models: Pareto-optimal Trade-offs for On-device AI,\" *Apple Machine Learning Research*, 2025. [18] A. Panangadan, \"Evaluating Accuracy of Generative AI for Diabetics\' Personalized Recipe Recommendations,\" *Proc. 2025 IEEE 26th International Conference on Information Reuse and Integration*, 2025.[51] [19] \"Effective Image Compression in Flutter: Reducing Latency for Multimodal AI APIs,\" *Ezility Medium Guide*, 2024.

Copyright

Copyright © 2026 Chinmay Mangela, Ibrahim Sheikh , Anas Sayyed, Juned Khan , Prof. Monika Pathare. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.