Food Classification using Deep Learning Techniques

Image recognition of real-world food plays an important role in the vast scope of applications. For example, food classification can estimate the gradients of any type of food to help diabetic and ill people to follow a healthy diet. Therefore, researchers have proposed methods and algorithms to recognize different types of food. However, the recognition process is a very challenging task due to the unlimited variations within the captured images of different kinds of food. Hence, we propose a new dataset which contains 20000 images of 15 different classes of local meals. A semi-automatic method is presented to provide the data annotations. The collected images are used to propose detector and recognizer models. The trained detector provides a detection rate of 81%. The train detector is designed by adopting YOLO v3-tiny. Additionally, the food classification process is performed based on two types of features: traditional features and learned ones. The results show the superiority of the deep learning features over the hand-craft features. Adopting the transfer learning techniques with the pre-trained model of Resnet50 achieves about 95% accuracy of classification rate.