The Sarong Paper: "Pattern Recognition of Sarong Fabric Using Machine Learning Approach Based on Computer Vision for Cultural Preservation” by Septiarini et al.

 

Objective

To develop a machine learning-based computer vision system for classifying Samarinda sarong patterns, which are part of Indonesia's traditional cultural heritage. Due to their visual similarity with sarongs from other regions, distinguishing these patterns manually is challenging.

---

Dataset

• Total images: 1000

• Classes: 6 (5 Samarinda types – Belang Hatta, Belang Negara, Belang Pengantin, Garanso, Kuningsau – and 1 non-Samarinda)

• Image capture: iPhone 6s camera, resized from 3024×4032 to 256×256 pixels

• Color spaces used: RGB, HSV, HSI, YIQ, YCbCr

---

Methodology

The classification pipeline involved:

1. Preprocessing: Image resizing and color space conversion

2. Feature Extraction:

   • Color: Extracted using Color Moments (mean, std, median, min, max) across five color spaces (total 75 features)

   • Texture:

     • GLCM (Gray Level Co-occurrence Matrix) – 32 features

     • LBP (Local Binary Pattern) – 10 features

   • Total features: 117

3. Feature Selection:

   • Correlation-Based Feature Selection (CFS) reduced features to 20

   • PCA reduced features to 48

4. Classification Models Tested:

   • Artificial Neural Network (ANN)

   • K-Nearest Neighbors (KNN)

   • Support Vector Machine (SVM)

   • Naive Bayes

   • Decision Tree

   • C.45 (a variant of Decision Tree)

---

Results

• Best classifier: ANN, which achieved 100% accuracy in all settings (with and without feature selection, across k-fold = 2, 5, 10)

• Other high-performing models: KNN, SVM

• CFS outperformed PCA in terms of feature reduction and performance

• Texture features (especially LBP and GLCM at 90°) improved model robustness when combined with color features

---

Comparative Advantage

• Outperformed prior works in batik and other textile classifications

• Used a larger dataset (1000 images)

• Achieved higher accuracy (100%) vs. 85–95% in prior studies

• Emphasized multi-color space extraction and minimal yet effective feature subset

---

Conclusion

The paper presents a robust and accurate method for classifying Samarinda sarongs using machine learning and image processing. The optimal combination of color moment + GLCM + LBP features, CFS-based selection, and ANN classifier ensures high classification accuracy and supports efforts in cultural heritage preservation.

See the complete Dissection Here