Comparison of Machine Learning Algorithms in Classifying Segmented Photographs of Food for Food Logging

Obesity is increasing globally and is a major cause for concern (WHO, 2016). The main cause of obesity is a result of a high calorie/ fat diet and when the energy is not burned off through exercise, then much of the excess energy will be stored as fat around the body. Obesity is a serious threat to an individual’s health as it can contribute to a range of major chronic conditions such as heart disease, diabetes, and some cancers (National Institutes of Health, 1998). Food logging is a popular dietary management method that has been used by individuals to monitor food intake. Food logging can include the use of text or images to document intake and research has shown that food intake monitoring can promote weight loss (Wing, 2001).There has been much research in using computer vision algorithms to classify images of food for food logging. Computer vision methods can offer a convenient way for the user to document energy intake. The motivation for this work is to inform the development of an application that would allow users to use a polygonal tool to draw around the food item for classification. This work explores the efficacy classifying segmented items of food instead of entire food images.This work explores machine learning (ML) techniques and feature extraction methods to classify 27 food categories with each category containing 100 segmented images. The image dataset used for this work comprises of 27 distinct food categories gathered from other research. (Jontou et al, 2009; Bossard et al, 2014). Non-food items contained in the images were removed to promote accurate feature selection (Figure 1).