A benchmark dataset to study the representation of food images

Abstract

It is well-known that people love food. However, an insane diet can cause problems in the general health of the people. Since health is strictly linked to the diet, advanced computer vision tools to recognize food images (e.g. acquired with mobile/wearable cameras), as well as their properties (e.g., calories), can help the diet monitoring by providing useful information to the experts (e.g., nutritionists) to assess the food intake of patients (e.g., to combat obesity). The food recognition is a challenging task since the food is intrinsically deformable and presents high variability in appearance. Image representation plays a fundamental role. To properly study the peculiarities of the image representation in the food application context, a benchmark dataset is needed. These facts motivate the work presented in this paper. In this work we introduce the UNICT-FD889 dataset. It is the first food image dataset composed by over 800 distinct plates of food which can be used as benchmark to design and compare representation models of food images. We exploit the UNICT-FD889 dataset for Near Duplicate Image Retrieval (NDIR) purposes by comparing three standard state-of-the-art image descriptors: Bag of Textons, PRICoLBP and SIFT. Results confirm that both textures and colors are fundamental properties in food representation. Moreover the experiments point out that the Bag of Textons representation obtained considering the color domain is more accurate than the other two approaches for NDIR.