Pengembangan Fitur Estimasi Gula Makanan Berbasis Image Recognition Menggunakan Machine Learning Pada Aplikasi Gulakupedia

Abstract

Type 2 Diabetes Mellitus is one of the most serious public health problems in Indonesia, with a national prevalence of 11.7% and an estimated 20.4 million adult patients. Controlling daily sugar intake is central to preventing and managing the disease, yet most traditional and home-cooked foods lack clear nutrition labels, making self-monitoring difficult. This study implements and integrates an image-recognition-based food sugar estimation feature into the Gulakupedia application, then measures the model accuracy and evaluates the feature functionality and usability. The problem is treated as a complex, multi-stage task: food classification from images, mapping sugar content to a standard serving size, and integration with the FatSecret nutrition database. A CRISP-DM framework was used to develop a YOLO object-detection model built on a Convolutional Neural Network (CNN) with transfer learning, while an Agile Scrum methodology guided the mobile application development. The detection model achieved a mAP50 of 0.691 on an independent test split, while the development target (mAP50 >= 0.70) was reached on the validation split (0.798); robustness against six image conditions was also examined. All 17 functional test cases in Black Box Testing passed (100% conformance), and the System Usability Scale (SUS) score was 75.97, categorized as Good/Acceptable. The quantized INT8 TFLite model (9.7 MB) ran on-device with a median end-to-end latency of about 2.04 s. The results indicate that the developed feature can help users monitor daily sugar intake independently and contribute to diabetes-complication prevention through practical, accessible technology.