The Motif Recognition Paper: Traditional Indian Textiles Classification using Deep Feature Fusion with Curvelet Transforms: Varshney et. al

 🔍 Objective

The paper proposes a hybrid feature fusion approach to classify traditional Indian textile patterns using:

• Pre-trained CNN models (InceptionResNetV2 and VGG16)

• Curvelet Transform (to capture curved edges better than wavelets)

---

📚 Dataset

• Total images: 1046

• Textile styles: Batik, Chikankari, Ikat, Kalamkari, Kashida, Madhubani, Warli

• Image size: 1200x1200

• Train/Test split: 90% / 10%

---

🔧 Methodology

1. Preprocessing: CLAHE (histogram equalization), resizing to 299x299 (IRV2) or 224x224 (VGG16).

2. Feature Extraction:

   • CNN-based: InceptionResNetV2 and VGG16

   • Handcrafted: Curvelet Transform (4 and 5 scale levels)

3. Fusion: Combine CNN and Curvelet features.

4. Classification: Using three models

   • XGBoost (XGB)

   • Gradient Boosting

   • Logistic Regression

5. Interpretability: Grad-CAM heatmaps to visualize feature focus areas.

---

📊 Key Results

• Best combination: Curvelet + InceptionResNetV2 without preprocessing

  • Accuracy: 97.15%

  • Precision: 98.24%

  • Recall/F1: 97.15%

  • Specificity: 99.52%

• VGG16 + Curvelet + Preprocessing achieved 91.43% accuracy.

---

🧠 Insights

• Curvelet transform adds rotational robustness and captures curved motifs better.

• InceptionResNetV2 with Curvelets performs significantly better than using CNN or Curvelet alone.

• Preprocessing has mixed impact depending on model used.

---

🔮 Conclusion

The fusion approach significantly boosts classification accuracy for Indian textile patterns. The method has potential applications in:

• Automated textile cataloging

• Design inspiration systems

• Cultural preservation via AI

Please refer to this link