Improvement Deep Learning Model for Batik Classification Used EfficientNetB2 and Augmentation Data
TL;DR
Findings indicate that EfficientNetB2 combined with data augmentation significantly improves classification accuracy, and the use of the EfficientNetB2 architecture does not require long computational time, making it highly practical and efficient.
Abstract
Batik is one of Indonesia’s cultural heritages that has been recognized by UNESCO. The diversity of batik motifs in Indonesia presents a challenge in classifying each motif from different regions. The purpose of this research is to improve batik motif classification using a deep learning model with the Convolutional Neural Network EfficientNetB2 architecture combined with data augmentation techniques. This study uses a dataset from Kaggle consisting of 1,964 images with seven batik motifs from the northern coast of Java: buketan, jlamprang, liong, mega mendung, negative, singa barong, and tujuh rupa. The augmentation technique applied in this study is traditional augmentation, including transformations such as horizontal flipping, rotation, and brightness adjustment. This was implemented to overcome dataset limitations, reduce the risk of overfitting, and increase dataset variation. The model integrates transfer learning through ImageNet pre-training to optimize the modification of the EfficientNetB2 architecture for specific batik motifs. EfficientNetB2 achieved an accuracy of 99% on the validation data, with precision, recall, and F1-scores consistently above 98%. This result is higher compared to previous research that applied the MPSO technique, which achieved 94% accuracy with precision, recall, and F1-scores of 78%. Model performance evaluation was conducted using a confusion matrix. Experimental results show that minimal misclassification occurred on the negative motif, with three samples incorrectly classified as jlamprang. These findings indicate that EfficientNetB2 combined with data augmentation significantly improves classification accuracy. Moreover, the use of the EfficientNetB2 architecture does not require long computational time, making it highly practical and efficient. This research contributes to the field of machine learning, the preservation of batik culture, and has the potential to support the batik industry in the commercial sector.