Indonesian Batik Pattern Recognition Using Siamese Network with Triplet Loss

This paper describes Indonesian batik pattern recognition depending on a Siamese Neural Network with a triplet loss function. The fabric art considered as Indonesian, that is batik is complex and comes in a wide variety of designs and hence hard to be staged and categorized. This paper has therefore developed a method that entails the acquisition and preprocessing of batik image dataset, data augmentation to enrich the dataset, the training of the Siamese Network coupled with CNNs (VGG16, DenseNet169, ResNet152) for feature extraction and optimizing the triplet loss function. This would show the efficiency of the Siamese Network in achieving high accuracy at a limited dataset over the traditional methods in pattern recognition. From several experiments carried out, the highest accuracy was obtained by the base model DenseNet169 with an accuracy level of 84% as a Siamese Network with trainable parameters set to false, where from this result there was an increase of 13% from the previous 71% when implemented on CNN with the same parameters. By carrying out this research, it is hoped that in the future more data can be collected from several regions in Indonesia regarding the diversity of batik patterns and motifs.