Visual food recognition using few-shot learning

• Main Author: Liu, Tianyi
• Other Authors: Yap Kim Hui
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Electrical and electronic engineering::Electronic systems::Signal processing
• Online Access: https://hdl.handle.net/10356/141132
• Institution: Nanyang Technological University

In recent years, smart food logging is becoming more popular. People can record their diet information and have better health management. Visual food recognition based on machine learning is one of the techniques to implement smart food logging system. It will extract visual features from food images and train a deep learning classifier for food recognition. However, deep learning algorithm is a data-driven approach which needs a large number of training data. Therefore, it will be a challenge when the training samples is not enough for some categories. In view of this, this project will study visual food recognition using few shot learning to solve the problem. We aim to study and use few-shot learning that can obtain the knowledge in the traning stage to recognize food categories with only a few samples. Some state-of-the-art few-shot learning networks include (1) Prototypical Network[1], (2) Relation Network[2], (3) Graph Neural network (GNN) Denoising Autoencoders[3] will be studied and evaluated in this project. Then we will apply them to do visual food recognition and evaluate their performance on a benchmark UECFOOD256 dataset.