The Food Recognition Benchmark: Using Deep Learning to Recognize Food in Images

Area of Science:

• Computer Vision
• Machine Learning
• Nutrition Informatics

Background:

• Automatic food recognition is crucial for applications like nutritional tracking.
• A public benchmark with non-biased data for reproducible algorithm development was lacking.
• The MyFoodRepo app provided a source for public food images from research cohorts.

Purpose of the Study:

• To establish a public benchmark for automatic food recognition.
• To create a diverse and representative dataset for algorithm training and evaluation.
• To foster open and reproducible research in food image analysis.

Main Methods:

• The MyFoodRepo-273 dataset was created, containing 24,119 images with 39,325 segmented polygons across 273 food classes.
• Four benchmark rounds were conducted, evaluating models on private test sets.
• Model performance was measured using mean average precision and mean average recall.

Main Results:

• Top models achieved a mean average precision of 0.568 (round 4) and mean average recall of 0.885 (round 3).
• The benchmark facilitated the deployment of high-performing models in the MyFoodRepo app.
• Experimental validation confirmed the results of the final benchmark round.

Conclusions:

• The established benchmark provides a valuable resource for advancing automatic food recognition.
• The benchmark setup is designed for future expansion in dataset size and diversity.
• This initiative supports the development of robust and reproducible food image analysis algorithms.