Selective Content Removal for Egocentric Wearable Camera in Nutritional Studies.

Area of Science:

• Biomedical Engineering
• Computer Science
• Nutritional Science

Background:

• Egocentric wearable cameras like Automatic Ingestion Monitor v2 (AIM-2) capture diet and eating behaviors.
• Image data from AIM-2 can reveal sensitive information, including bystanders and private documents.
• Existing methods for privacy protection in wearable camera data are insufficient.

Purpose of the Study:

• To propose and evaluate a novel approach for automatic image redaction to protect privacy in egocentric camera data.
• To address bystander and context privacy concerns associated with wearable cameras used in nutritional studies.

Main Methods:

• Utilized semantic segmentation with a deep learning neural network for selective content removal.
• Developed an automatic image redaction method for privacy protection.
• Applied the method to images captured by the Automatic Ingestion Monitor v2 (AIM-2).

Main Results:

• Achieved high precision (0.87) and recall (0.94) for bystander privacy removal.
• Reported excellent precision and recall (0.97 and 0.98, respectively) for context privacy removal.
• Demonstrated the effectiveness of deep learning for selective content removal in egocentric imagery.

Conclusions:

• Selective content removal using deep learning is a highly effective method for addressing privacy concerns in egocentric wearable cameras.
• The proposed approach offers a desirable solution for enhancing privacy in nutritional studies utilizing wearable camera technology.
• Automatic image redaction significantly improves the privacy protection capabilities of dietary monitoring systems.