Automatic Diet Recording Based on Deep Learning

Diet assessment plays an important role in the prevention and treatment of many diseases, but it takes too much time and money to do a thorough long-term diet record. In order to make the recording efficient and affordable, we aim at building an automatic diet recording system, which can improve the automation level of diet recording through deep learning algorithms. And two contributions have been made in this paper. First, several deep learning models are trained and evaluated on the commonly used food image datasets, and after fusing several popular convolution neural network models, the top-1 accuracy on UEC FOOD 100, UEC FOOD 256, Food-101 and ChineseFoodNet datasets can reach 84.0%, 75.9%, 84.5% and 79.8% respectively. Second, based on the trained deep learning models, we build an automatic diet recording prototype system, which performs well during the test in a real canteen. The work of this paper suggests that deep learning can achieve a high accuracy in the automatic identification of dietary objects, which will help to promote the research in the field of disease prevention and rehabilitation.