Obesity Detection in Thermal Imaging Using Convolution Neural Network: A Comparison with Machine Learning Models

Abstract

The purpose of the proposed study is to classify the thermal images of abdomen, forearm, and shank regions in obesity and normal subjects using the Speeded Up Robust Feature (SURF)-based machine learning classifiers compared to the convolutional neural network. Total populations of 60 subjects, among which 30 normal and 30 obese subjects, were recruited for the proposed study. The thermal images were acquired in abdomen, forearm, and shank regions using FLIR thermal camera (FLIR A300). SURF-based feature extraction method is used along with SVM classifiers for binary classification of adult obesity and normal subjects. After data augmentation process, convolution neural network VGG19 is used for the classification of obesity and normal subjects, and its performance is compared with SVM classifiers. The SURF-based SVM classifier provides an accuracy of 93% compared to VGG19 net. The machine learning classifier provided better results compared to the pre-trained models in CNN in the thermal image obesity analysis.KeywordsThermal imagingSpeeded up robust featureSupport vector machineConvolutional neural network