Date of Award

Summer 8-19-2022

Level of Access Assigned by Author

Open-Access Thesis

Degree Name

Master of Science (MS)

Department

Spatial Information Science and Engineering

Advisor

Nimesha Ranasinghe

Second Committee Member

Nicholas Giudice

Third Committee Member

Silvia Nittel

Abstract

Food is a vital component of our everyday lives closely related to our health, well-being, and human behavior. The recent advancements of Spatial Computing technologies, particularly in Human-Food interactive (HFI) technologies have enabled novel eating and drinking experiences, including digital dietary assessments, augmented flavors, and virtual and augmented dining experiences. When designing novel HFI technologies, it is essential to recognize different food and beverages and their internal attributes (i.e., food sensing), such as volume and ingredients. As a result, contemporary research employs image analysis techniques to identify food items, notably in digital dietary assessments. These techniques, often combined with AI algorithms, analyze digital food images to extract various information about food items and quantities. However, these visual food analyzing methods are ineffective when: 1) identifying food’s internal attributes, 2) discriminating visually similar food and beverages, and 3) seamlessly integrating with people’s natural interactions while consuming food (e.g., automatically detecting the food when using a spoon to eat). This thesis presents a novel approach to digitally recognize beverages and their attributes, an essential step towards facilitating novel human-food interactions. The proposed technology has an electrical impedance measurement unit and a recognition method based on deep learning techniques. The electrical impedance measurement unit consists of the following components: 1) a 3D printed module with electrodes that can be attached to a paper cup, 2) an impedance analyzer to perform Electrical Impedance Spectroscopy (EIS) across two electrodes to acquire measurements such as a beverage’s real part of impedances, imaginary part of impedances, phase angles, and 3) a control module to configure the impedance analyzer and send measurements to a computer that has the deep learning framework to conduct the analysis. Two types of multi-task learning models (hard parameter sharing multi-task network and multi-task network cascade) and their variations (with principal component analysis and different combinations of features) were employed to develop a proof-of-concept prototype to recognize eight different beverage types with various volume levels and sugar concentrations: two types of black tea (LiptonTM and TwiningsTM English-Breakfast), two types of coffee (StarbucksTM dark roasted and medium roasted), and four types of soda (regular and diet coca-cola, and regular and diet Pepsi). Measurements were acquired from these beverages while changing volume levels and sugar concentrations to construct training and test datasets. Both types of networks were trained using the training dataset while validated with the test dataset. Results show that the multi-task network cascades outperformed the hard parameter sharing multi-task networks in discriminating against a limited number of drinks (accuracy = 96.32%), volumes (root mean square error = 13.74ml), and sugar content (root mean square error = 7.99gdm􀀀3). Future work will extend this approach to include additional beverage types and their attributes to improve the robustness and performance of the system and develop a methodology to recognize solid foods with their attributes. The findings of this thesis will contribute to enable a new avenue for human-food interactive technology developments, such as automatic food journaling, virtual flavors, and wearable devices for non-invasive quality assessment.

Download

Files over 10MB may be slow to open. For best results, right-click and select "save as..."

Share