Comparative Performance Analysis of Xception, DenseNet201, and VGG16 Deep Learning Models with Ensemble Techniques for Spice Classification

Spices are an essential part of Indonesian society, serving as important flavor enhancers and valuable natural resources. Accurate identification is challenging because spices share similar textures, colors, and shapes. Furthermore, the use of instant spices has reduced public familiarity with traditional spice varieties. To address these challenges, integrating deep learning techniques into computer vision offers a promising approach for fast, accurate spice classification. This study uses transfer learning, augmented with hyperparameter optimization using the Weighted Average Ensemble method, to improve performance with a limited dataset. Experiments used three advanced transfer learning models: Xception, DenseNet201, and VGG16, along with five optimizers: Adam, RMSprop, SGD, AdaMax, and Nadam. This methodology focused on classifying 12 different types of spices. Results showed that two specific ensemble combinations achieved optimal performance. The combination of Xception and DenseNet201 achieved accuracies of 97.08% and 90.09%, respectively, while the combination of Xception, DenseNet201, and VGG16 achieved accuracies of 97.42% and 89.82%, respectively. These results demonstrate the effectiveness of ensemble learning for spice classification and its potential for automating processes in the spice industry. Future research should expand the dataset and prioritize model generalization to further advance practical applications.