Diet Application That Can Recognize Turkish Foods With Deep Learning

Abstract

The problem of manually entering the meals one by one, which is one of the main problems in many traditional diet applications, was solved with the help of the proposed deep learning models integrated into the mobile application. Two different models resulting from the deep learning study were used in the application. The first model detects the food in the environment with real-time object detection and the second one recognizes the type of the food in the detected plate. A data set containing 102 different types of food belonging to Turkish cuisine and approximately 500 photographs of each type of food was collected and used in the training of the second model. The proposed TurkishFoodNet network with the number of three, five, seven, nine, eleven and thirteen layers were examined in the training of both models. Apart from this network, training operation were also held with Tensorflow Lite Image Classifier and MobilNetV2. According to the test results, Tensorflow Lite Image Classifier and TurkishFoodNet_L11 gives the highest accuracy for both detecting food and recognizing the type of food with an accuracy of 93% and 84%, respectively.