A new deep learning-based food recognition system for mobile terminal

Abstract

With the improvement of people's health awareness, people pay more attention to their health. In recent years, the intelligent health management system based on food recognition technology has become popular, which can help users maintain healthy eating habits. However, applying the current deep learning method in mobile phones and other terminal devices is difficult, mainly because the terminal devices have the low computing power and the network needs to perform many calculations during operation. In this paper, we have adopted the methods of parameter reconstruction and calculation graph fusion to reduce the network computing load so that it can run in real-time in terminal devices, and the detection speed on Snapdragon 778G SOC exceeds 7 FPS. Besides, experiments on the VIPER-FoodNet (VFN) dataset show that our model has a high mean average precision (mAP) of 9.17% compared with the current advanced model.