USE OF COMPUTER VISION TECHNIQUES FOR AUTOMATIC FOOD CLASSIFICATION BY SIZE

Background and objectives: Computer vision (VC) techniques are a novel tool for the food and nutrition area. The classification of foods according to size using VC techniques would allow the estimation of portion sizes and grams, optimize the evaluators time and reduce subjectivity, but research in this area is still incipient. Therefore, we set out to develop an automatic and reproducible method of food classification by size. Methods: To evaluate the method we used oranges, which are much consumed and are available in the local market. We measured the dimensions (greater diameter) of the oranges with a caliber and weighed each one of them, later we classified them in: small, medium and large. We created a closed environment to obtain high quality images, and developed a program that takes the obtained images as input data and we applied VC techniques which includes image processing, segmentation, extraction of characteristics and classification using the algorithm KNN. Finally we estimated the sensitivity and specificity of the results obtained through the program. Results: We obtained photographs of 207 oranges (124 used for training of the program (60%) and 83 for the corresponding tests of the same (40%)), the average greater diameter according to the measurements of the expert was 69.1 ± 4.4 millimeters and the obtained from the program was 68.9 ± 4.1. The Pearson correlation showed r = 0.68 (p 0.05). After 100 training with random assignment of oranges in the ratio of 60% to 40%, the program had a sensitivity and specificity equal to 1.0. Conclusions: The proposed method could provide assistance to nutrition professionals in the process of estimating the size of foods and consequently portions and grams. The database of images and physical measurements generated would enable more studies in the area.