A multimodal deep learning approach for food tray recognition

Abstract

[en] Food recognition, object detection and classification applied to the food domain, is the main topic of this work. We have studied the problem of recognising food instances in tray images of self-service restaurants and have proposed a novel multimodal deep learning approach. From images and daily menus, the model presented uses two state of the art models in object detection and classification and a multimodal neural network to make significantly refined predictions compared to the baseline object detection model, achieving a class weighted average F1-score of 0.862. An ensemble model built from the proposed and the baseline models, also presented in this work, improves the results achieving a class weighted average F1-score of 0.877.