Enhancing Small Dataset Performance: Data Augmentation and Transfer Learning in Indonesian Traditional Foods Classification

One of the difficulties in deep learning is making a model with small dataset. Overfitting, outliers, low accuracy, noise, and general poor performance often caused by lack of information given to the model in learning process. Transfer learning is a method where solving a problem will be carried out using a model that has previously been trained by other. With transfer learning, models that have been trained can be used to continue training on smaller or specific datasets. Besides transfer learning, data augmentation can also be used as alternative solution. By images modification, augmenation process will give additional images to the model. In this research, several methods will be carried out, which are augmentation and transfer learning to produce a model that is able to recognize images of Indonesian traditional food. Based on the test results, it was found that the use of data augmentation can reduce the accuracy gap between training and validation data. With 89% of accuracy and decrease in loss values for all data (training, validation, and testing data), it proved that the utilization of data augmentation and transfer learning enhanced overall model performance.