Validating an automated image identification process of a passive image-assisted dietary assessment method: proof of concept

Abstract

AbstractObjective:To validate an automated food image identification system, DietCam, which has not been validated, in identifying foods with different shapes and complexities from passively taken digital images.Design:Participants wore Sony SmartEyeglass that automatically took three images per second, while two meals containing four foods, representing regular- (i.e., cookies) and irregular-shaped (i.e., chips) foods and single (i.e., grapes) and complex (i.e., chicken and rice) foods, were consumed. Non-blurry images from the meals’ first 5 min were coded by human raters and compared with DietCam results. Comparisons produced four outcomes: true positive (rater/DietCam reports yes for food), false positive (rater reports no food; DietCam reports food), true negative (rater/DietCam reports no food) or false negative (rater reports food; DietCam reports no food).Setting:Laboratory meal.Participants:Thirty men and women (25·1 ± 6·6 years, 22·7 ± 1·6 kg/m2, 46·7 % White).Results:Identification accuracy was 81·2 and 79·7 % in meals A and B, respectively (food and non-food images) and 78·7 and 77·5 % in meals A and B, respectively (food images only). For food images only, no effect of food shape or complexity was found. When different types of images, such as 100 % food in the image and on the plate, <100 % food in the image and on the plate and food not on the plate, were analysed separately, images with food on the plate had a slightly higher accuracy.Conclusions:DietCam shows promise in automated food image identification, and DietCam is most accurate when images show food on the plate.