Processing of Egocentric Camera Images from a Wearable Food Intake Sensor
Viprav B. Raju, Edward E. Sazonov · 2019
The cameras in passive food intake sensors are worn on the body and thus are prone to motion blur that may reduce visibility of the food items in the image. Additionally, captured images may contain personal information, such as faces of other people, making it imperative to uphold the privacy of the wearer and the people appearing in the images. This paper proposes an automatic image sharpness estimation followed by face removal. A new metric for image sharpness was defined and a classifier for binary classification of sharp and blurred images was tested. Then, face detection was followed by face removal to enable privacy. The methods were developed and tested on two datasets consisting of a total of 1600 images each. The datasets were inclusive of the several lighting conditions and image capture situations, such as indoor/outdoor, low-light/normal lighting conditions, food/non-food images, and stationary/moving images. Validation showed that the misclassification rate for the binary classification of sharpness of 10% to 15%. The accuracy of human face detection was 86.67% with an F1 score of 0.81.