Enhanced Semarang batik classification using deep learning: a comparative study of CNN architectures

Abstract

Batik is an important part of Indonesia’s cultural heritage, with each region producing unique designs. In Central Java, Semarang is known for its distinctive batik patterns that reflect rich local traditions. However, many people are still unfamiliar with these designs, which threatens their preservation. This study develops an automated system to classify Semarang batik patterns, showing how technology can help safeguard cultural heritage. A convolutional neural network (CNN) approach was used to recognize ten batik types, including Asem Arang, Asem Sinom, Asem Warak, Blekok, Blekok Warak, Gambang Semarangan, and Kembang Sepatu. Pre-processing steps—such as image resizing, cropping, flipping, and rotation—improved model performance and reduced complexity. Five CNN architectures (MobileNetV2, ResNet-50, DenseNet-121, VGG-16, and EfficientNetB4) were tested using 224×224 input size, Adam optimizer, ReLU activation, and categorical cross-entropy loss. Results show VGG-16, ResNet-50, and DenseNet-121 achieved perfect accuracy (1.0) on a dataset of 3,000 locally collected images. These findings highlight CNN models’ strong potential for batik pattern recognition, supporting digital preservation of Indonesian culture.