Transfer Learning Implementation for Batik Besurek Motifs Classification: A Comparative Study

This research discussion will be a comparative model proposal for introducing Besurek batik fabric by implementing MobileNet, ResNet50, InceptionV3 and VGG16. This research also analyzed the influence of augmentation data methods, namely rotation, zoom, width shift and height shift, on the performance of the Besurek batik fabric recognition algorithm based on accuracy, precision and recall results. The size of the batik besurek fabric motif image as an input image is $224 \times 224$ with three channels consisting of red, green, and blue, totalling five classes, namely kaligrafi, bunga rafflesia, burung kuau, lung paku and rambutan. The ratio of sharing training data and test data is 70:30. Accuracy evaluations are carried out on several epochs and data that have been shared. Based on the experiment, the accuracy value of MobileNetV2 can be obtained by training accuracy of $\mathbf{98.14 \%}$, validation accuracy of $\mathbf{94.60 \%}$, and testing accuracy of $\mathbf{96.46 \%}$. Then, the accuracy value of ResNet50 can be obtained by training accuracy at $\mathbf{99.60 \%}$, validation accuracy at 97.44, and testing accuracy at $\mathbf{98.12 \%}$. The accuracy value of InceptionV3 can be obtained by training accuracy at $96.35 \%$, validation accuracy at $96.59 \%$, and testing accuracy at $\mathbf{92.50 \%}$. Finally, the accuracy value of VGG16 can be obtained by training accuracy at $\mathbf{99.67 \%}$, validation accuracy at $96.02 \%$, and testing accuracy at $\mathbf{95.00 \%}$. The accuracy of VGG 16 can be concluded as the best model for Batik recognition.