A Python implementation of rotation-invariant texture classification using Local Binary Patterns (LBP) and its variants, achieving 100% accuracy on the Brodatz rotated texture dataset.

Example: Bark texture at different rotation angles from the Brodatz dataset

This project implements a robust texture classification system that is invariant to rotation transformations. The system uses Local Binary Patterns (LBP), a powerful texture descriptor, combined with multi-resolution analysis and local variance features.

Key Achievement: 100% classification accuracy on 13 Brodatz textures across 7 rotation angles (0°, 30°, 60°, 90°, 120°, 150°, 200°).

For detailed methodology and complete analysis, please refer to:

• Implementation Code - Complete source code
• Project Presentation - Detailed slides with methodology and references

• Multiple LBP Variants:

  • Basic LBP with circular neighborhoods
  • Rotation-invariant LBP (LBPri)
  • Rotation-invariant uniform LBP (LBPriu2)

• Multi-Resolution Analysis:

  • Multiple (P, R) configurations: (8,1), (16,2), (24,3)
  • Combined multi-scale operators

• Variance Descriptors:

  • Local variance (VAR) computation
  • Joint LBP/VAR histograms for enhanced discrimination

• Efficient Implementation:

  • Vectorized NumPy operations
  • Pre-computed lookup tables for rotation invariance
  • Custom bilinear interpolation for circular neighborhoods

1. Multi-resolution improves performance: Larger neighborhoods (P=16, R=2 or P=24, R=3) capture coarser spatial structures
2. VAR is highly effective: Adding local variance complements LBP by encoding contrast information
3. Perfect classification achieved: LBP₁₆,₂/VAR achieves 100% accuracy across all training scenarios

The project uses 13 textures from the USC-SIPI Brodatz rotated texture database:

• Textures: bark, brick, bubbles, grass, leather, pigskin, raffia, sand, straw, water, weave, wood, wool
• Rotations: 0°, 30°, 60°, 90°, 120°, 150°, 200° (7 angles per texture)
• Format: Grayscale, 512×512 pixels, 8-bit TIFF
• Total images: 91 (13 textures × 7 angles)

rotate/
├── bark.000.tiff
├── bark.030.tiff
├── bark.060.tiff
...
├── wool.200.tiff

• Training: Use one rotation angle per texture (e.g., 0°)
• Testing: Evaluate on remaining 6 rotation angles
• Patch extraction: 16×16 subimages from 512×512 images
• Cross-validation: Train on 0°, 30°, 60°, 90° separately

rotation-invariant-texture-classification/
│
├── 22b2505_finalProject_CS663.py  # Main implementation
├── README.md                       # This file
│
├── docs/
│   ├── 22b2505_finalProject_CS663.pdf  # Detailed presentation
│   └── images/                     # Documentation images
│
├── rotate/                         # Dataset directory (not included)
│   ├── bark.000.tiff
│   └── ...
│
└── 22b2505_results/               # Generated results (after running)
    ├── lbp_maps/
    ├── histograms/
    ├── confusion_matrices/
    └── experiment_summary.txt

1. Lookup Tables (LUTs):

   • Pre-computed rotation-invariant mappings
   • RIU2 pattern mappings
   • Avoids expensive per-pixel recomputation

2. Vectorized Operations:

   • NumPy broadcasting for neighbor comparisons
   • Batch processing of all pixels simultaneously

3. Bilinear Interpolation:

   • Custom implementation for circular neighborhoods
   • Handles non-integer pixel coordinates efficiently

4. Variance Quantization:

   • Percentile-based cut-points from training data
   • Consistent binning across train/test sets

5. Numerical Stability:

   • Laplace smoothing for histogram models
   • Log-probabilities to prevent underflow

• LBPOperator: Core LBP computation with configurable P, R
• TextureClassifier: Multi-operator training and classification
• Helper functions: Image loading, patch extraction, visualization

1. T. Ojala, M. Pietikäinen and T. Mäenpää, "Multiresolution Gray-scale and Rotation Invariant Texture Classification with Local Binary Patterns," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, no. 7, pp. 971-987, 2002. IEEE Xplore

2. Brodatz Texture Database, USC-SIPI Image Database. Link