Traditional Classification Based On Deep Learning Case Studies Of Hodgepodge And Other Foods

Abstract

This study aims to classify the image of traditional Indonesian food with a high level of visual similarity using the Deep Learning method based on the Residual Network (ResNet) architecture. The dataset consists of seven food classes, namely gado-gado, gudeg, nasi padang, rawon, rendang padang, satay, and chicken soup. The preprocessing stage is carried out through image size normalization, format conversion, and data augmentation in the form of rotation, flipping, and lighting adjustment to increase data diversity and reduce the risk of overfitting. The experiment was conducted using two variants of the ResNet architecture, namely ResNet-18 and ResNet-34, with an image size of 224×224 pixels and a batch size of 32. The training results showed that ResNet-18 achieved a training accuracy of 97.59% and a validation of 83.00% in the 30th epoch, while ResNet-34 obtained a training accuracy of 95.08% and a validation of 89.29% in the 50th epoch. Based on the results of the evaluation through the confusion matrix, the model was able to recognize most classes well, although there were still errors in classes that had similar colors and textures such as rendang padang and rawon. Models with deeper layers have been shown to have better generalization capabilities to complex image variations. This research proves that the ResNet architecture is effectively used in the classification of traditional Indonesian food images and has the potential to be applied to artificial intelligence (AI)-based automatic recognition systems in the fields of culinary, tourism, and digital education. Further development may include the application of advanced transfer learning, the integration of models into mobile or web applications, as well as the use of Explainable AI (XAI) methods such as Grad-CAM or SHAP to improve the interpretability of classification results.