Efficient Deep Learning Methods for Food Localization in Canteen Trays

This paper presents a comparative study of various efficient state-of-the-art image segmentation models applied to the challenging task of food localization in trays in canteen environments. Using the UNIMIB2016 dataset, which comprises images of canteen trays with multiple food items, we evaluate the performance of ten deep learning-based methods in terms of their segmentation accuracy measured by the Jaccard Index, and computational efficiency measured by Multiply-Accumulate (MACs) operations and the number of parameters. Our results illustrate a tradeoff between computational demand and accuracy, with DABNet achieving the highest accuracy but at a cost of lower efficiency compared to others, while models like ENet or EDANet offer a balanced solution suitable for real-time applications. The study not only benchmarks these models but also discusses the implications of different architectural choices, such as the use of dilated and depth-wise convolutions, which influence the models' performance. This work aims to guide the selection of appropriate segmentation models for dietary management systems in canteen settings, contributing to advancements in automated food service operations and dietary monitoring.