Date of Award
Fall 12-9-2024
Degree Type
Thesis
Degree Name
Master of Science in Computer Science
Department
CCSE
Committee Chair/First Advisor
Maria Valero
Second Advisor
Taeyeong Choi
Third Advisor
Kazi Aminul Islam
Abstract
Effective diet monitoring is crucial for various health conditions, including diabetes management, as it helps individuals maintain optimal nutrition while controlling blood sugar levels. This study investigates the potential of Large Language Models (LLMs) to automate food logging by estimating nutritional content from food images. Using 100 food images, the models’ calorie estimations were compared against ground truth values from provided by a dietician. A smartphone app was developed to log meals via photos, processed by an LLM API to fetch detailed macronutrient data. The analysis shows that all models have reasonable accuracy, but performance is affected by outliers (errors >50%). Removing these outliers improves model performance significantly. Post-outlier removal, Gemini 1.5 Flash had the highest R-squared value (0.50), indicating a better fit and variance capture in calorie estimations.