Deep learning approach in food detection : An application for nutrition tracking

Abstract

Tracking food intake can give insights into eating habits, hence it is useful to confront the rising public health threat of obesity and overweight. Because of high pace of modern life, people usually do not have time for manually recording their daily meals. Therefore, it is necessary to have an application which can accelerate the tracking process by automatically detecting foods from input image, collecting nutrition facts and recording. Deep learning methods have already proved promising in the food detection challenge. Among these methods, TensorFlow Object Detection API is an open-source framework, consists of many pre-trained models that can be used to train and deploy a custom object detection model. Furthermore, there are some pre-trained models uses the MobileNet which is an efficient convolutional neural network for mobile vision applications.

The project uses aforementioned framework for training with modified UECFOOD-256 dataset, which is reduced from 256 classes in original dataset to 10 classes. The obtained model can detect and localize 10 types of food present in image. After that, the model is used in web application which is developed responsive for both mobile and desktop devices. Its aims are detecting food object, summarizing total nutrition, recording to database, and providing insight for users.