Deep Learning Framework for Food Item Recognition and Nutrition Assessment

The increasing prevalence of dietary-related health conditions necessitates the development of precise, automated nutritional analysis systems leveraging advancements in artificial intelligence. This paper presents a novel approach to food image analysis using a YOLOv8 object detection model and a custom nutritional assessment, to provide detailed nutritional information from food images. The system ingests user-uploaded food images and employs advanced preprocessing pipelines to standardize the input. The YOLOv8 model, implemented via ONNX Runtime, identifies food items and delineates them in bounding boxes with high confidence. To address dataset het-erogeneity, novel algorithms for class merging, underrepresented class removal, and dynamic class weighting were devised, culmi-nating in a refined dataset of 95,000 instances across 214 food cat-egories. Nutritional values, including macronutrient breakdowns, are derived through a custom volumetric computation function, correlating plate areas occupied by detected items with database-derived density metrics. Experimental evaluations reveal superior model performance, achieving mean Average Precision (mAP) scores of 0.7941 at IoU=0.5. Qualitative analyses further validate the system's efficacy under diverse real-world conditions such as overlapping objects and inconsistent lighting. Deployed on Heroku as a Flask web application, this system exhibits computational efficiency and scalability, supporting real-time dietary monitoring through mobile devices. The proposed system serves as a proof-of-concept that can be improved to provide a robust, scientifically grounded tool for health management by bridging gaps between AI capabilities and practical nutritional needs, with potential extensions to broader healthcare applications.