Indonesian Food Image Recognition Using Convolutional Neural Network

Abstract

Food image recognition becomes more interesting because it can be useful in health industry to obtain many information from a food image such as calorie, nutrition, carbohydrates, fats and protein. The challenging part of food image recognition is to recognize a food image with different background, intensity, and perspective. Convolutional Neural Network (CNN) seems to be the right choice to build a powerful model that able to recognize food image accurately. Current researches in food recognition use American fast food and Japanese food as the dataset. Different dataset has different treatment to obtain good result. Color and presentation has important role as features. Therefore, this research proposes to recognize five kinds of popular Indonesian food such as meatball (bakso), grilled chicken (ayam bakar), satay (sate), gado-gado (mixed vegetables with peanut sauce), and rendang using Convolutional Neural Network approach. Convolutional Neural Network with standard architecture and inception-v4 model are chosen as techniques with Adam as the optimizer function. Experimental results show the implementation of CNN for image recognition can achieve the top-1 testing accuracy around 76.3% for standard network and 95.2% for inception-v4 network.