Indonesian Food Image Classification with Grad-CAM Visualization using High Performance Computing DGX A100
Lingga Rohadyan, Eri Prasetyo Wibowo, Setia Wirawan, Ana Kurniawati · 2024
In recent years, the importance of nutrition has gained significant attention in Indonesia, particularly with initiatives aimed at providing nutritious meals to children. Image classification technology offers a promising solution, potentially allowing individuals to quickly identify foods and estimate their nutritional value using just a food image. In this paper, we present an approach to classifying Indonesian food images using convolutional neural network, with a focus on increasing model interpretability using visualization techniques. We used the EfficientNetV2 architecture, taking advantage of its state-of-the-art performance and computational efficiency. Our dataset consisted of 2,000 images, categorized into ten distinct classes which were pre-processed and utilized to fine-tune the pre-trained EfficientNetV2 model. Training was carried out using the NVIDIA DGX A100 machine, significantly reducing the training time. To improve the transparency of the classification process, we used Gradient-weighted Class Activation Mapping (Grad-CAM) to identify the regions of interest in the images that had the greatest influence on the model’s predictions. This helped us better understand the model’s decision-making process and suggest areas for improvement. Our experiments achieved an accuracy of 87.25% within 50 epochs, demonstrating the model’s capability to correctly classify Indonesian food images. We also discovered areas of misclassification and used Grad-CAM visualizations to determine the causes, which will guide future dataset and model enhancements.