Fine Tuning CNN Pre-trained Model Based on Thermal Imaging for Obesity Early Detection

Abstract

Obesity is a complex disease that causes serious impact health, such as diabetes mellitus, cardiovascular disease, cancer, and stroke. An early obesity diagnosis/ detection method is required to prevent the increasing number of obese people. This study aims to: (i) fine-tune the pre-trained Convolutional Neural Network (CNN) models to build an early detection of obesity and (ii) evaluate the model performance in terms of classifying performance, computation speed, and learning performance. The thermal images acquisition procedure was conducted with 18 normal subjects and 15 obese subjects to build a thermal images dataset of obesity. Pre-trained CNN models: VGG19, MobileNet, ResNet152V, and DenseNet201 were modified and trained using the acquired dataset as the input. The training results show that the DenseNet201 model outperformed other models regarding classifying accuracy: 83.33 % and learning performances. At the same time, the MobileNet model outperformed other models in terms of computation speed with training elapsed time: 12 seconds/epoch. The proposed DenseNet201 model was suitable for implementation as an early screening system of obesity for health workers or physicians. Meanwhile, the proposed MobileNet model was suitable for mobile applications' early detection/diagnosis of obesity.