<![CDATA[Previous research on traditional food image classification demonstrated that the VGG16 architecture could achieve good classification performance. However, the study focused on a single architecture and did not provide a comprehensive comparison between predictive performance and computational efficiency across multiple models. This limitation highlights the need to evaluate other architectures in order to identify a model that is not only accurate but also suitable for different deployment environments. This study aims to compare the performance and efficiency of three transfer learning architectures, namely VGG16, ResNet50, and MobileNetV2, for traditional food image classification. The dataset consisted of 2,445 images grouped into ten classes: batun bedil, bubur injin, jaje lukis, jaje piling, jaje wajik, kaliadrem, klepon, laklak, ongol-ongol, and pisang rai. The dataset was divided using a 70:15:15 ratio into 1,706 training images, 363 validation images, and 376 testing images. All models were trained under the same configuration, using an input size of 224 × 224 pixels, batch size of 64, Adam optimizer, learning rate of 0.0001, categorical crossentropy loss, data augmentation, and 10 epochs. Model evaluation was conducted using accuracy, precision, recall, F1-score, training time, model size, and inference time per image. The results showed that ResNet50 achieved the best classification performance, with an accuracy of 84.04%, precision of 85.15%, recall of 85.16%, and F1-score of 84.89%. MobileNetV2 achieved an accuracy of 82.45% with the smallest model size of 12.98 MB, while VGG16 obtained an accuracy of 73.67%. These findings indicate that ResNet50 is more suitable for systems that prioritize classification performance, whereas MobileNetV2 is more appropriate for applications requiring a lightweight model. The contribution of this study lies in the comparative evaluation of performance and efficiency across three architectures under the same experimental setting, resulting in practical model recommendations based on deployment requirements.]]>
