FoodLog

Abstract

With the increase of the number of food images on the Internet, we have been developing a food-logging system which has an automated analysis function as a Web application. It can distinguish food images from other images, analyze the food balance, and visualize the log. In this paper, we demonstrate how the performance can be improved by the personalized models. Because our Web application has an interface to review and correct the food analysis results, the generation of the personalized models can be done on-line. Experimental results using two hundred images showed that the extracted image feature vectors differ from user to user but on the other hand the feature vectors and the food balance of each user have a strong correlation. Therefore, the accuracy of the food balance estimation was improved from 37% to 42% on average by the personalized classifier.