Klasifikasi citra makanan/non makanan menggunakan metode Transfer Learning dengan model Residual Network

Abstract

People's activities that often upload photos of food before eating are common in popular social media, such as Facebook, Instagram, Twitter, and Pinterest. The food image data circulating on social media can be used for business purposes, such as analyzing customer behavior patterns. However, not all of the uploaded images are food images, so that before the analysis is carried out, it is necessary to do a classification task between food images and non-food images. Therefore, the researcher proposes automatic food image classification using transfer learning method using the Residual Network model version of 18 (ResNet-18). Residual Network model is used because it has a residual connection mechanism to solve the vanishing gradient problem. In addition, transfer learning was chosen because this method leverages the features and weights that have been generated in the previous training process on large and more general data (Imagenet) and thus reduce computation time and increase accuracy. The test was carried out by comparing the capabilities of the ResNet18 model with AlexNet. In addition, the fine tuning and freeze layer methods used to improve the quality of the model were also carried out in this study. In the experiment, the data set was divided into 3,000 images for training data and 1,000 images for test data, while the evaluation used was correctness accuracy. The results obtained in the ResNet18 model, namely the fine tuning training method, produced an accuracy value of 0.981 while the freeze layer resulted in the best accuracy value of 0.988. The AlexNet model that uses the fine tuning training method produces an accuracy value of 0.970 while the freeze layer produces the best accuracy value of 0.978. It can be concluded that the mechanism with the best accuracy is found in the RestNet18 architecture using the freeze layer 1-3 with an accuracy of 0.988.