Food Components Recognition from Still Images Using Multi-Label Learning

Abstract

Food recognition, a recent research area in image processing, helps identify food items to keep track of the food
consumed, thereby maintaining a healthy diet. However, the task of food recognition is challenging due to the deformable nature of food items. Usually, there are more than one food item in a meal making the task more challenging. Therefore, the aim of this work is to develop a deep learning model to detect and enumerate visual food components present in a meal. In the multi-label learning approach, food images were collected to build a food image dataset, which comprised 2150 images. The images were pre-processed. Contrast Limited Adaptive Histogram Equalization was then applied followed by scaling to fit as input into the model for training/testing. Thereafter, Deep (VGG-16) and Dense (DenseNet50) models were used to extract deep features. The final layer of the model was applied with a multi-label technique to train on the selected features. The multi-label model was tested using appropriate metrics in which VGG-16 performed better than DenseNet50 with an accuracy of 91.90%, hamming loss of 8.10%, loss of 0.26%, precision of 73.49%. An independent test set was used on the model which showed impressive results. It was observed from this study that the proposed approach performed excellently well in predicting Nigerian Food components. It is recommended that this work be applied in real world this work in real world scenario such as dietary tracking to monitor food intake. Human-Computer Interaction with automatic purchasing systems at restaurants can be used to speed up services.