Region-based convolutional network-based image processing model for determining the amount of carbohydrates in food

Type 1 Diabetes Mellitus (T1DM) is an autoimmune disease that increases the number of patients every year around the world. This disease, which affects nutritional methods and carbohydrate counting is important, requires individuals to constantly monitor themselves and keep a record of their nutrition. The ability of people to easily calculate the ingredients and nutritional values of their meals with computer vision methods has been made possible by studies conducted in recent years. Studies on the calculation of the volume of the food with both 2-dimensional images and 3-dimensional images have been done and are being done. With the acceleration of developments on artificial neural networks in recent years, studies on food recognition and volume estimation have started to achieve high success. In this study, we estimate the amount of carbohydrate in the meals with the volume information we have obtained through calculations that we have made by inputting the diameters of the plates after training with the Detectron2 system. Our system was found to be accurate with an average error of around 7 percent, with an error margin ranging from 1 percent to 15 percent. Although the change in error rates is related to the way the food is served, the estimation of such a high success rate with a single image shows that our system is successful. The system is suitable to be developed for 3D volume estimation with visuals obtained from different angles, and is ready to be developed and used with more classes and training images.