Calorie Estimation of Daily Meals using Machine Learning

Abstract

Eating a healthy, well-balanced diet is becoming more and more important as nutritionrelated diseases rise in prevalence worldwide. Over the past few years, recording and calculating daily calorie intake has also become essential in maintaining health. To maintain a proper healthy diet, systems that track user activities are becoming common practice through Dietary Assessment Systems. However, calculating calories is still a challenging task. It was traditionally calculated by manually logging in the details of daily intake, which is cumbersome for most users. Recent research has tried to overcome these challenges and calculate the calories of foods using various techniques but these works still have limitations, including limited dataset of images of local dishes in public datasets. While automatic systems are being developed and machine learning is replacing manual traditional systems, this thesis provides a simple approach for calculating the calories of any food through images without the need to manually log daily caloric intake. Convolutional Neural Networks (CNNs) were therefore used in this investigation since they have produced ground-breaking results in image classification and for calculating calories concept of OpenCV was utilized. Furthermore, a new custom dataset of Pakistani food images was developed through web-scraping. This thesis also explores and assesses performances of three different state-of-the-art approaches to classify images using deep learning techniques. Initially, for image classification CNN was used followed by 03 x Transfer Learning models namely XceptionNet, VGGNet, and MobileNet. The results obtained show overall accuracy of CNN with 79.82% and XceptionNet with an accuracy of 84.69%. Furthermore, food’s calories were estimated using volumetric approach. Lastly, this study can also be utilized to develop an application that calculates the calories of any food based on images.