Class-Incremental Learning for Enhanced Food Recognition

Yen Thi Hong Tran, Phap-Minh Nguyen, Hung-Phuoc Nguyen, Khanh-Duy Nguyen

Abstract: While many deep learning CNN models have been utilized for multiclass food dataset classification, traditional food image classification methods struggle with generalization and often require retraining when new cuisines or additional classes are introduced. This challenge is especially pronounced for under-represented cuisines, such as Vietnamese food. In this paper, we examine the impact of Class Incremental Learning (CIL) on food datasets, with a particular focus on Vietnamese food datasets. We apply CIL and compare its performance with traditional classification methods across various datasets, including 30VNFoods, VinaFood21, Food101, FoodX251, and ISIA Food-500, considering both accuracy and efficiency. Our experiments demonstrate strong stability from MEMO (68-75% accuracy on old classes), though Replay and ICaRL outperform MEMO in terms of new class accuracy. Furthermore, the initial choice of a global food dataset significantly influences the accuracy of recognizing Vietnamese food classes in the incremental training phase. The UCEFOOD100 dataset achieves the highest accuracy in both scenarios: 0 initial classes with 20 incremental classes (82.13%) and 20 initial classes with 30 incremental classes (87.68%).

External IDs:dblp:conf/iccais/TranNNN24