Analisis Pengaruh Augmentasi Data Terhadap Performa Transfer Learning MobileNetV2 dalam Klasifikasi Citra Makanan Indonesia

Abstract

Limited training data and high visual variability in Indonesian food images often cause deep learning–based classification models to experience overfitting and difficulties in accurately recognizing new images. To address this issue, this study applies eight data augmentation scenarios to a transfer learning–based MobileNetV2 model for classifying 10 Indonesian food categories, namely Ayam Pop, Bakso, Gado-Gado, Mie Goreng, Nasi Goreng, Rawon, Rendang, Sate, Soto, and Telur Balado. The dataset consists of 500 images used for training, which are divided into 70% training data and 30% validation data, along with 100 additional images used as an independent test set. The applied augmentation techniques include rotation, zoom, brightness adjustment, contrast adjustment, photometric (brightness + contrast), geometric (rotation + zoom), and a combined scenario integrating all augmentation techniques, as well as a baseline scenario without augmentation. Model performance was evaluated using accuracy, precision, recall, F1-score, and confusion matrix. The results indicate that all augmentation techniques improve the model performance compared to the baseline scenario, which only achieved 80.00% validation accuracy and showed signs of overfitting. The rotation scenario achieved the best performance with a validation accuracy of 91.87% and an independent test accuracy of 87.00%. These findings demonstrate that appropriate data augmentation can improve both the accuracy and generalization capability of the MobileNetV2 model in Indonesian food image classification under limited data conditions.