The Identification of Red-Meat Types using The Fine-Tuned Vision Transformer and MobileNet Models

Abstract

For reasons related to poverty or lack of quality control over food in some countries of the world, there is still food adulteration. Low-cost meats such as donkey or pork are marketed as lamb or beef. This is morally dangerous but may be more dangerous for some people who are allergic to certain types of meat or who have religious reservations. With the rapid development of artificial intelligence techniques, it is possible to build a model capable of differentiating between different types of meat. This study aims to build a model capable of differentiating between different types of red meat. It also aims to compare performance between the very state of art CNN in computer vision with the transformer architecture. For this goal, a limited dataset from an online repository was obtained. The dataset contains RGB images of beef, horse, and pork meats. The images were processed, and various data augmentation techniques were applied. Then vision transformer ViT and mobile net models with and without fine-tuning were built. To measure the models' behavior, several performance evaluation criteria were applied. The best testing accuracy is 97% achieved by the fine-tuned ViT model. This study showed the effectiveness of applying the transformer architecture and especially the fine-tuned ViT model in the areas of image classification even on a limited dataset.