Food Image Recognition and Nutrition Estimation via Deep Learning

Abstract

This thesis proposes a large-scale food image dataset namely AIFood dataset and a system for food category recognition of food image using deep learning and nutrition estimation In order to build a food image dataset for food image recognition we collect the food images from existing food image datasets and a food website and relabel all images using 24 food categories In addition we preprocess the images using automatic white balancing and contrast limited adaptive histogram equalization to improve the visual quality of images We calculate the information of the image and set constraints to detect if the image is needed to be preprocessed Next for food image recognition we modify a 50-layer residual convolutional neural network (ResNet50) by removing the maximum pooling to decrease the loss the information and replacing part of convolutional layers with dilated convolution to increase receptive field of convolutional layers After training the neural network using AIFood dataset we can achieve 83 63% and 76 90% performance of Micro-F1 and Macro-F1 for food category prediction respectively The performance of our modified ResNet50 is better than the original ResNet50 Next we collect the nutrition information from the Ministry of Health and Welfare Taiwan and calculate the nutrition of each category per one person meal Finally the nutrition content in the food image is the sum of the nutrition of all food categories detected by the neural network in the food image

Date of Award	2020
Original language	English
Supervisor	Gwo-Giun Lee (Supervisor)