Batik Motif Classification Using Deep Learning and Transfer Learning

Table of Contents

• Introduction
• Dataset Overview
• Methodology
• Data Preprocessing
• Model
• Results and Discussion
  • Model Performance
  • Key Insights
  • Limitations
• Conclusion
• License
• Contributors

Introduction

This project focuses on classifying three popular Indonesian batik motifs—Parang, Mega Mendung, and Kawung—using deep learning techniques, specifically Convolutional Neural Networks (CNNs) and Transfer Learning. These motifs are well-known in Indonesian culture, and this automated classification system aims to assist in identifying these motifs efficiently.

While the dataset contains a variety of batik motifs, this project specifically focuses on Batik Parang, Batik Mega Mendung, and Batik Kawung, leveraging the power of deep learning to classify these motifs with high accuracy.

Dataset Overview

The dataset for this classification project is sourced from Kaggle - Indonesian Batik Motifs, which includes various batik motifs. For this project, we focus on three iconic motifs:

1. Batik Parang: 50 images
2. Batik Mega Mendung: 46 images
3. Batik Kawung: 45 images

While the full dataset contains 983 images, this project narrows the focus to these three motifs to demonstrate deep learning techniques in image classification.

Methodology

Data Preprocessing

The data preprocessing steps are essential to ensure the images are properly prepared for deep learning models:

1. Image Resizing: All images were resized to 224x224 pixels, as required by most deep learning architectures.
2. Data Augmentation: Techniques such as rotation, zooming, and horizontal flipping were applied to increase the diversity of the dataset, preventing overfitting.
3. Normalization: Pixel values were normalized to range between 0 and 1.
4. Encoding: The target labels (Parang, Mega Mendung, and Kawung) were one-hot encoded for multi-class classification.

---

Model

Transfer Learning

Several deep learning models were trained to classify the three batik motifs:

• From Scratch: Custom CNN models were built and trained from scratch with and without data augmentation.
• Transfer Learning: Pre-trained models such as VGG16, InceptionV3, ResNet50, and MobileNetV2 were used for transfer learning. These models, pre-trained on large datasets like ImageNet, were fine-tuned on the batik dataset to adapt the learned features for motif classification tasks.

Among these, MobileNetV2 achieved the highest accuracy and was the model of choice for this classification.

Model Training and Evaluation

The models were trained using the Adam optimizer and categorical cross-entropy loss function. Early stopping was applied to prevent overfitting, and validation data was used throughout the training to monitor the model's performance.

---

Results and Discussion

Model Performance

The transfer learning approach, particularly using MobileNetV2, significantly outperformed custom CNN models trained from scratch. MobileNetV2, a highly efficient model, was fine-tuned to classify the three batik motifs, achieving 93.75% accuracy on the validation set. This demonstrates the power of transfer learning in handling smaller, specialized datasets.

Key Insights

1. The Power of Transfer Learning: Transfer learning, especially using MobileNetV2, provided the best performance compared to custom models trained from scratch.
2. Image Augmentation: Data augmentation was crucial in preventing overfitting, allowing the model to generalize better to unseen data.
3. High Accuracy for Batik Kawung: The model showed the highest accuracy in recognizing Batik Kawung, indicating that the visual patterns of this motif were clearer for the model to distinguish.

Limitations

1. Limited Dataset: The relatively small dataset of 983 images may have limited the model's ability to generalize fully.
2. Overfitting: Despite using transfer learning, some overfitting signs appeared, particularly when distinguishing Batik Mega Mendung from the other motifs.
3. Similar Motifs: Certain motifs share visual features that made classification more challenging, especially in differentiating between Batik Mega Mendung and Batik Parang.

Conclusion

This project demonstrates the effectiveness of deep learning and transfer learning for classifying Indonesian Batik motifs. MobileNetV2 was the model of choice due to its high efficiency and 93.75% accuracy on the validation set. The results show that transfer learning with pre-trained models provides significant improvements, especially when working with small datasets. Future work can focus on expanding the dataset, improving model fine-tuning, and incorporating additional batik motifs to enhance model accuracy.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contributors

• Steve Marcello Liem
• Davin Edbert Santoso Halim
• Felicia Andrea Tandoko