Libraries at University of Nebraska-Lincoln
Date of this Version
11-24-2020
Abstract
Purpose- Estimation of food portions is necessary in image based dietary monitoring techniques. The purpose of this systematic survey is to identify peer reviewed literature in image-based food volume estimation methods in Scopus, Web of Science and IEEE database. It further analyzes bibliometric survey of image-based food volume estimation methods with 3D reconstruction and deep learning techniques.
Design/methodology/approach- Scopus, Web of Science and IEEE citation databases are used to gather the data. Using advanced keyword search and PRISMA approach, relevant papers were extracted, selected and analyzed. The bibliographic data of the articles published in the journals over the past twenty years were extracted. A deeper analysis was performed using bibliometric indicators and applications with Microsoft Excel and VOS viewer. A comparative analysis of the most cited works in deep learning and 3D reconstruction methods is performed.
Findings: This review summarizes the results from the extracted literature. It traces research directions in the food volume estimation methods. Bibliometric analysis and PRISMA search results suggest a broader taxonomy of the image-based methods to estimate food volume in dietary management systems and projects. Deep learning and 3D reconstruction methods show better accuracy in the estimations over other approaches. The work also discusses importance of diverse and robust image datasets for training accurate learning models in food volume estimation.
Practical implications- Bibliometric analysis and systematic review gives insights to researchers, dieticians and practitioners with the research trends in estimation of food portions and their accuracy. It also discusses the challenges in building food volume estimator model using deep learning and opens new research directions.
Originality/value- This study represents an overview of the research in the food volume estimation methods using deep learning and 3D reconstruction methods using works from 1995 to 2020. The findings present the five different popular methods which have been used in the image based food volume estimation and also shows the research trends with the emerging 3D reconstruction and deep learning methodologies. Additionally, the work emphasizes the challenges in the use of these approaches and need of developing more diverse, benchmark image data sets for food volume estimation including raw food, cooked food in all states and served with different containers.