Multiple Food or Non-Food Detection in Single Tray Box Image using Fraction of Pixel Segmentation for Developing Smart Nutrition Box Prototype

Abstract

Smart Nutrition Box is a hardware prototype to predict food leftover measurement as well as the nutrition of the leftover. In the previous approach, there was a need of trained observer to conduct the analysis. Human observer may produce subjective judgement, so that algorithm which is embedded in a prototype is proposed to get rid of the bias. Black background of tray box is used, and two menus are served in this paper. The problem on raw dataset of images is reflection and it affects the result of segmentation, since it is considered to determine the leftover measurement precisely. In this paper, we focus on how to classify image of food and non-food image in each compartment of tray box by using pixel segmentation before going to further stage of prediction. Automatic cropping is applied by means of rectangle contour detection for each compartment. Combination of L of HSL and V color channel of HSV color spaces are utilized to remove glare in each compartment. The ratio of segmented pixel is a fraction of detected object and the area of compartment. There are 10 out of 12 of tray box images containing multiple food that are correctly classified as food and non-food. The accuracy reaches 95.83% in all compartments using luminosity (L) 50% of lower upper white masking and 100% of upper white masking. It is proved that fraction pixel segmentation is sufficient to be embedded as one of features in Smart Nutrition Box prototype.