Deep Learning-based Algorithms for Malaysian Food Image Recognition and Calories Estimation

Diabetes is a prevalent health concern among Malaysian adults, with dietary intake monitoring playing a crucial role in managing the disease. Existing dietary tracking approaches often overlook the diversity of Malaysian cuisine and therefore suffer from low accuracy in recognizing and estimating portion sizes for local dishes. To address these limitations, this study employs deep learning techniques for food recognition to develop an effective calorie estimation system tailored to Malaysian food culture. The study focuses on the investigation of three distinct models which are MobileNetV2, VGG19, and ResNet-50 in accurately identifying and classifying Malaysian food items from images. At present, there is lack of food dataset images for Malaysian cuisine, which presents a significant challenge to calorie counter or food logging app developers. Thus, 20 classes of Malaysian cuisine were generated with a total of 2,076 images used in the models. We then compared the performance of the three models using accuracy, precision, recall and F1-score metrics. The MobileNetV2 model shows a promising result with an accuracy of 97.83% compared to VGG19 with 71.05% accuracy. Meanwhile ResNet-50 achieved the lowest accuracy of 18.11%. Additionally, we evaluated the performance of MobileNetV2 on larger and more diverse cuisines to measure its scalability.