Implementation of Convolutional Neural Networks to Recognize Images of Common Indonesian Food

Abstract

One of the applications of computer vision in the popular culture is food recognition which is popularized in the internet with a "hotdog and not hotdog" problem. Food recognition is also useful in many popular lifestyle apps such as calorie counter app or any diet related app. In this paper is proposed a CNN aided technique for recognizing food that is common in Indonesia. The technique is consist of 3 main phases, one is pre-processing normalize the data one is model formation and training which is known as the common binary classifier template, boosted with pooling and evaluated by cross-entropy technique in this paper the model used is the pre-trained model to test the testing data afterward, which is show a promising result with a relatively short training time. The experiments focused on how CNN can be used as a component to recognize food so that in the future it can be used to develop better calorie counter applications. In this experiment 10,000 data were used for training and 50 tests for each food category with a total of 500 food image data used for testing with the best accuracy reaching 88% for one of the categories.