Rizki Ripai, Fajar Mahardika, Fazar Sidik, Nurul Badriah, Angga Maulana Purba

Deep Learning-Based Classification of Cikadu Batik Motifs Using ResNet50 and MobileNetV2

Introduction

Deep learning-based classification of cikadu batik motifs using resnet50 and mobilenetv2. Classify Cikadu Batik motifs using deep learning (ResNet50, MobileNetV2). Preserves cultural heritage via accurate Indonesian batik recognition & digitization, comparing model efficiency.

Abstract

Batik motif recognition is essential for cultural heritage preservation and the digitization of traditional Indonesian textile knowledge. This study proposes a deep learning-based framework for the automatic classification of Cikadu Batik motifs from Tanjung Lesung, Banten — a regionally distinct batik pattern that has not been systematically studied in prior computational literature. Two convolutional neural network (CNN) architectures were implemented and comparatively evaluated under identical experimental conditions: ResNet50, a high-capacity model employing residual skip connections, and MobileNetV2, a lightweight model utilizing depthwise separable convolutions and inverted residual blocks. A curated dataset of 2,500 images spanning five motif classes was constructed through collaboration with local batik artisans, preprocessed via resizing (224×224), pixel normalization, and augmentation (rotation, zoom, horizontal flip, brightness adjustment), and partitioned using a stratified 70:15:15 split. Both models were trained with transfer learning from ImageNet weights, using the Adam optimizer (lr=0.0001), categorical cross-entropy loss, batch size of 32, and early stopping over 30 epochs. Model evaluation employed accuracy, precision, recall, F1-score, AUC-ROC, inference time, and parameter count. ResNet50 achieved 95.51% accuracy, 95.67% precision, 95.34% recall, 95.50% F1-score, and 99.56% AUC-ROC, with an inference time of 18.2 ms and 25.64 million parameters. MobileNetV2 achieved 92.13% accuracy, 92.28% precision, 91.98% recall, 92.13% F1-score, and 98.89% AUC-ROC, with an inference time of 8.7 ms and only 3.54 million parameters — approximately 7× lighter and 2× faster. These results empirically establish a clear accuracy-efficiency trade-off, with ResNet50 favored for accuracy-critical server-based systems and MobileNetV2 better suited for real-time mobile deployment. This study constitutes the first published benchmark for deep learning-based Cikadu Batik classification and provides a principled basis for architecture selection in regional batik recognition applications

Full Text

Comments

You need to be logged in to post a comment.