Research Article
A Deep Learning-Based Dessert Recognition System for Automated Dietary Assessment
@INPROCEEDINGS{10.1007/978-3-031-06368-8_4, author={Dimitrios-Marios Exarchou and Anastasios Alexiadis and Andreas Triantafyllidis and Dimosthenis Ioannidis and Konstantinos Votis and Dimitrios Tzovaras}, title={A Deep Learning-Based Dessert Recognition System for Automated Dietary Assessment}, proceedings={Wireless Mobile Communication and Healthcare. 10th EAI International Conference, MobiHealth 2021, Virtual Event, November 13--14, 2021, Proceedings}, proceedings_a={MOBIHEALTH}, year={2022}, month={6}, keywords={CNN Computer vision Deep learning Image recognition Dessert recognition Food Image Image pre-processing Diabetes Obesity}, doi={10.1007/978-3-031-06368-8_4} }- Dimitrios-Marios Exarchou
Anastasios Alexiadis
Andreas Triantafyllidis
Dimosthenis Ioannidis
Konstantinos Votis
Dimitrios Tzovaras
Year: 2022
A Deep Learning-Based Dessert Recognition System for Automated Dietary Assessment
MOBIHEALTH
Springer
DOI: 10.1007/978-3-031-06368-8_4
Abstract
Over the past few years, a significant part of scientific research has been focused on the assistance of patients who suffer from obesity or diabetes. Monitoring the food intake through self-report in diet control applications has been proven both time-consuming and non-practical and can be easily sidelined especially by children. In this paper, we propose the design and development of a novel system, which will assist obese or diabetic patients. We have implemented transfer learning as well as fine-tuning to different pre-trained CNN models to automatically distinguish dessert from non-dessert food images. For further training of these deep neural networks, a new dataset was constructed, which derived from the original Food-101 dataset. To be precise, 19 categories of desserts were used, which correspond to 19K images combined with 19K images of non-desserts. Google InceptionV3 architecture appeared to have the best performance, reaching a validation accuracy of 95.89%. To demonstrate feasibility of out platform and the independence of data biases, we constructed another data collection of food images, which was captured under challenging light and angle of capture conditions.
