The AI4Food-NutritionDB is the first food image database that incorporates a detailed nutrition taxonomy. The database includes:
The database was constructed by combining 7 existing food image datasets, resulting in over 560,000 food images.
The authors propose a standardized experimental protocol and benchmark for evaluating food recognition systems based on the nutrition taxonomy. They assess the performance of state-of-the-art deep learning models (Xception and EfficientNetV2) on both intra-database and inter-database scenarios.
The intra-database results show that the models achieve high accuracy in categorizing food at the coarser levels (e.g., 82.04% Top-1 accuracy for 19 food categories), but the performance decreases as the granularity increases (66.28% Top-1 accuracy for 893 food products).
The inter-database evaluation demonstrates that the deep learning models pre-trained on the AI4Food-NutritionDB significantly outperform models trained only on ImageNet when applied to the challenging VireoFood-251 database, achieving up to 88.80% Top-1 accuracy.
The AI4Food-NutritionDB and the associated benchmark are publicly available, fostering the development of advanced food computing technologies for personalized nutrition.
To Another Language
from source content
arxiv.org
Deeper Inquiries