Food Image Classification with Deep Features

Abstract

In this paper, deep feature extraction, feature concatenation and support vector machine (SVM) classifier are used for efficient classification of food images. Classification of foods according to their images becomes a popular research task for various reasons such as food image retrieval and image based self-dietary assessment. For deep feature extraction, pre-trained AlexNet and VGG16 models are considered. The features of size 4096 are extracted from fc6 and fc7 layers and concatenated with various combinations to determine best deep feature sequence for food image classification. The concatenated features are then classified with SVM. Three publicly available datasets namely FOOD-5K, FOOD-11 and FOOD-101 are used in evaluation of the proposed method and the accuracy metric is considered for performance evaluation. The experimental results show an accuracy of 99.00% for FOOD-5K dataset and 88.08% and 62.44% for FOOD-11 and FOOD-101 datasets, respectively. We further carried out experiments with fine-tuning of a pre-trained CNN model on FOOD-101 dataset and obtained 79.86% accuracy score. The obtained results are also compared with some other methods and it is seen that our performance is better than the other methods on FOOD-11 and FOOD-101 datasets.