Automatic food detection in egocentric images using artificial intelligence technology.

Area of Science:

• Computer Science
• Nutrition Science
• Biomedical Engineering

Background:

• Dietary assessment is crucial for understanding human health and lifestyle.
• Egocentric wearable cameras offer a novel method for capturing real-world dietary data.
• Manual analysis of large image datasets for dietary assessment is time-consuming and raises privacy concerns.

Purpose of the Study:

• To develop an artificial intelligence (AI)-based algorithm for automatic food item detection from egocentric wearable camera images.
• To facilitate automated dietary assessment and reduce the burden of manual data processing.
• To address privacy concerns associated with collecting and analyzing personal dietary images.

Main Methods:

• Acquisition of large egocentric image datasets (eButton datasets 1 and 2) from free-living individuals using a wearable device.
• Manual selection and classification of images containing real-world activities, including food-related and non-food-related scenarios.
• Development and application of a convolutional neural network (CNN) for automated food/non-food image classification and detection.

Main Results:

• Cross-dataset testing on eButton dataset 1 achieved high food detection accuracy (91.5% and 86.4%) with half-data training/testing.
• For eButton dataset 2, the AI model demonstrated 74.0% sensitivity and 87.0% specificity when considering both 'food' and 'drink' as food images.
• When considering only 'food' items, the model achieved 85.0% sensitivity and 85.8% specificity.

Conclusions:

• AI technology can reliably detect foods from low-quality, egocentric images captured by wearable cameras.
• The developed algorithm offers a practical solution for automated dietary assessment, significantly reducing data processing efforts.
• This approach enhances user privacy by automating the identification of food items, minimizing the need for manual review of sensitive images.