🇧🇩 BanglaDressNet: Fine-Grained Classification of Traditional Bangladeshi Dress

Official repository for the conference paper: "Multi-Level Ensemble Learning for Fine-Grained Classification of Traditional Bangladeshi Dress Using Deep Transfer Models"

Accepted at: 2025 7th International Conference on Electrical Information and Communication Technology (EICT) 18–20 December 2025, Khulna University of Engineering & Technology, Khulna-9203, Bangladesh

---

📄 Abstract

The fine-grained classification of traditional costumes is challenging due to subtle differences between categories and significant internal variations. This paper introduces BanglaDressNet, an 8,000-image dataset of 16 categories of Bangladeshi dresses (e.g., Saree, Panjabi, Lungi) collected through web scraping and filtered with perceptual hashing.

We evaluated 10 ImageNet-pretrained CNNs via transfer learning. The best three models—DenseNet121, EfficientNetB1, and Xception—achieved test accuracies of 95.75%, 95.58%, and 94.75%, respectively. To enhance performance, we proposed three ensemble strategies: validation accuracy-weighted averaging, stacked logistic regression, and stacked MLP. The Weighted Averaging Ensemble achieved the highest performance with 96.75% test accuracy, 96.75% F1-score, and 99.91% macro ROC-AUC.

---

🔗 Quick Links

Resource	Link
📄 Paper	Awaiting Publication
🐍 Kaggle Dataset
🤗 Hugging Face

---

💾 Dataset Description

BanglaDressNet is a fine-grained dataset for traditional Bangladeshi dress classification. (Novel Dataset)

• Total Images: 8,000
• Classes: 16 Traditional Dress Categories
• Split: 70% Train, 15% Validation, 15% Test
• Preprocessing: Resized to 224x224, Normalized

Class List

1. Blouse
2. Dhoti
3. Dupatta
4. Fatua
5. Frock
6. Gamcha
7. Koti
8. Lungi
9. Maxi
10. Muffler
11. Panjabi
12. Petticoat
13. Salwar-Kameez
14. Sando-Genji
15. Saree
16. Shawl

---

🧠 Methodology

We employed a Multi-Level Ensemble Learning approach to capture diverse features from different architectures.

Models Implemented

• DenseNet121 (Texture features)
• EfficientNetB1 (Multi-scale features)
• Xception (Shape features)
• Ensemble Strategies:
  • Weighted Averaging
  • Stacked Logistic Regression
  • Stacked MLP

---

📊 Result Analysis

Our proposed Weighted Averaging Ensemble outperformed individual models and other ensemble techniques.

Model	Test Accuracy	F1-Score	ROC-AUC
Weighted Avg. Ensemble	96.75%	0.9675	0.9991
Stacking LR	96.58%	0.9659	0.9980
Stacking MLP	96.50%	0.9651	0.9970
DenseNet121	95.75%	0.9579	0.9989
EfficientNetB1	95.58%	0.9558	0.9991
Xception	94.75%	0.9472	0.9981

---

📝 Citation

If you use this dataset or code in your research, please cite our paper:

@misc{md__musfiqur_rahman_2025,
	title={BanglaDressNet},
	url={https://www.kaggle.com/dsv/13932271},
	DOI={10.34740/KAGGLE/DSV/13932271},
	publisher={Kaggle},
	author={Md. Musfiqur Rahman},
	year={2025}
}

---

👥 Authors

• Md. Musfiqur Rahman
• Sudipa Biswas
• S.N.M. Rayhan
• Md Shohan Mia
• Md Mahbubur Rahman Tusher

---

📬 Contact

For any questions, please open an issue in this repository or contact the authors.