A comparative study of deep learning methods for food classification with images

Abstract

The monitoring of food intake plays a key role in avoiding different health problems such as diseases, but also difficulties linked with mental and physical wellbeing. Food monitoring consists firstly of, identifying food types and secondly, recommending the appropriate food for better nutrition. This study addresses the aspect of identifying and classifying food types using Artificial Intelligence (AI) through computer vision, deep learning and transfer learning techniques. We used a dataset containing over 16000 food images, classified into 11 categories: Bread, Dairy product, Dessert, Egg, Fried food, Meat, Noodles/Pasta, Rice, Seafood, Soup, and Vegetable/Fruit. On average, an additional 15884 images were computationally generated for each experiment with data augmentation techniques such as rotation, shifting, shearing, zooming and flipping. Our experiments with different Convolutional Neural Networks (CNN) demonstrate that transfer learning with the EfficientNetV2 model achieved a significant validation accuracy of 94.5 % in classifying the different types of food from images. Also, data augmentation contributes to deal with the imbalanced dataset, with the model achieving an F1-score of 94.7 %. Together with drop out and early stopping techniques, data augmentation also contributes to keep the model safe from overfitting.