A Comparison of Several Approaches for Image Recognition used in Food Recommendation System

摘要

In food recommendation systems, users can use mobile devices to capture images of the dishes they eat. The types of dishes in images will be automatically recognized and input to the recommendation system to suggest other dishes which the users likely to enjoy. Food image recognition is therefore an essential part of the food recommendation system. This used to be a hard problem in computer vision as many foods are very similar in color and texture. Thanks to the development of artificial intelligent and especially deep learning techniques, it is much easier to build a program to recognize the type of foods in the image. This paper examines several approaches, from traditional machine learning to state-of-the-art deep learning techniques for food image recognition to provide a comparison of the performance of these techniques. To this end, a new dataset of Vietnamese cuisine including 12,017 photos of 15 dishes has been built to test algorithms. Traditional machine learning techniques including Histogram of Gradient (HOG) and Scale-Invariant Feature Transform (SIFT) and state-of-the-art deep learning models including VGG16, MobileNet, ANN, Resnet18, Resnet50, Densenet121 have been used for extracting features in the food images. Logistic Regression (SF) and SoftMax (SM) classification have been used for classification using the extracted features. Based on the comparison results provided in this paper, one can choose appropriate techniques for image recognition to build a good food recommendation system.