Eliminate the hardware: Mobile terminals-oriented food recognition and weight estimation system

Area of Science:

• Computer Vision
• Medical Nutrition
• Digital Health

Background:

• Food recognition and weight estimation are crucial for digital nutrition therapy and health detection.
• Deep learning advancements have expanded image-based recognition across various fields.
• Existing methods often require costly or burdensome wearable equipment.

Purpose of the Study:

• To develop a non-wearable food recognition and weight estimation system (nWFWS) for smartphones.
• To assist patients and physicians in monitoring diet-related health conditions.
• To simplify automatic food intake estimation using mobile phone photography.

Main Methods:

• Utilized deep convolutional neural networks and a visual-inertial system for image pixel collection.
• Trained the system on 612 high-resolution food images with diverse traits.
• Developed a relationship model between food pixel area and measured weight.

Main Results:

• Achieved an 89.60% accuracy rate in identifying 1,455 food pictures.
• Successfully determined the weight of untested food images with high correlation between predicted and actual values.
• Demonstrated feasibility and relative accuracy for automated dietary monitoring.

Conclusions:

• The developed non-wearable food recognition and weight estimation system (nWFWS) is a feasible and accurate tool.
• The system offers a cost-effective and simplified approach to dietary monitoring and nutritional assessment.
• Mobile-based image analysis holds significant potential for advancing personalized health management.