Towards personalised training of machine learning algorithms for food image classification using a smartphone camera

Abstract

This work is related to the development of a personalised machine learning algorithm that is able to classify food images for food logging. The algorithm would be personalised as it would allow users to decided what food items the model will be able to classify. This novel concept introduces the idea of promoting dietary monitoring through classifying food images for food logging by personalising a machine learning algorithm. The food image classification algorithm will be trained based on specific types of foods decided by the user (most popular foods, food types e.g. vegetarian). This would mean that the classification algorithm would not have to be trained using a wide variety of foods which may lead to low accuracy rate but only a small number of foods chosen by the user. To test the concept, a range of experiments were completed using 30 different food types. Each food category contained 100 images. To train a classification algorithm, features were extracted from each food type, features such as SURF, LAB colour features, SFTA, and Local Binary Patterns were used. A number of classification algorithms were used in these experiments; Nave Bayes, SMO, Neural Networks, and Random Forest. The highest accuracy achieved in this work was 69.43% accuracy using Bag-of-Features (BoF) Colour, BoF-SURF, SFTA, and LBP using a Neural Network.