Improving Batik Pattern Classification using CNN with Advanced Augmentation and Oversampling on Imbalanced Dataset

Abstract

In image classification task, imbalanced dataset is a problem that often occurs. Batik pattern data also suffers this problem, mainly because of the poor quality of available images and rarity of certain patterns. In this research, we employed a novel ad- vanced augmentation and oversampling techniques on the imbalanced dataset to address this issue. This approach enhanced the diversity of the images, encompassing variations in color, contrast, wrinkles, and warps that may be present in batik garments. We employed two CNN models, DenseNet169 and VGG-16, along with three different training methods for our study. These methods included training without oversampling and advanced augmentation, training with oversampling, and training with both oversampling and advanced augmentation. The results showed that the best accuracy was achieved with DenseNet169 with our oversampled and augmented dataset, with an accuracy of 84.62%. Additionally, VGG-16 also performed well on said dataset, achieving an accuracy of 82.56%.Our results suggested that by using our oversampling & advanced augmentation on the dataset,the model performance improved compared to plain data and oversampled data.