Multi-View Clustering Research: Feature Extraction and DBSCAN Analysis

📋 Project Overview

This repository contains a comprehensive research project on multi-view clustering using various feature extraction methods and DBSCAN clustering algorithm. The research focuses on analyzing the effectiveness of different feature combinations for clustering tasks on two benchmark datasets: COIL-20 and Caltech-7.

🎯 Research Objectives

• Primary Goal: Investigate the performance of multi-view clustering using different feature extraction methods
• Secondary Goal: Compare clustering algorithms (K-Means vs DBSCAN) and optimize parameters
• Research Question: Which combination of features provides the best clustering performance for image datasets?

📊 Datasets Used

1. Caltech-7 Dataset (Primary Focus)

• Source: Subset of Caltech-101 dataset
• Classes: 7 categories with varying image counts
• Total Images: 1,474 images
• Classes Breakdown:
  • Faces: 435 images
  • Motorbikes: 798 images
  • Dollar-Bill: 52 images
  • Garfield: 34 images
  • Snoopy: 35 images
  • Stop-Sign: 64 images
  • Windsor-Chair: 56 images

2. COIL-20 Dataset (Secondary Analysis)

• Source: Columbia Object Image Library
• Cars Subset: 144 images (objects 10 and 14)
• Full Dataset: 1,440 images (20 objects × 72 rotations each)
• Purpose: Validation and comparative analysis

🔧 Feature Extraction Methods (Views)

Our research implements 6 different feature extraction techniques to create multiple views of the data:

1. LBP (Local Binary Pattern)

• Purpose: Texture analysis and local pattern recognition
• Parameters:
  • Radius: 3
  • Number of points: 24
  • Method: 'uniform'
• Feature Dimension: 26 features
• Advantage: Robust to illumination changes, captures local texture patterns

2. HOG (Histogram of Oriented Gradients)

• Purpose: Shape and edge detection
• Parameters:
  • Pixels per cell: (8,8)
  • Cells per block: (2,2)
  • Orientations: 9
• Feature Dimension: 8,100 features
• Advantage: Captures shape information and spatial relationships

3. WM (Wavelet Moments)

• Purpose: Multi-resolution analysis
• Method: Haar wavelet decomposition
• Feature Dimension: 8 features (mean & std of 4 coefficients)
• Advantage: Captures both frequency and spatial information

4. CENTRIST (Census Transform)

• Purpose: Texture description using census transform
• Implementation: LBP-based with 'uniform' method
• Feature Dimension: 26 features
• Advantage: Computationally efficient, good for texture classification

5. Gabor Features

• Purpose: Edge and texture detection at different orientations
• Feature Dimension: 48 features
• Advantage: Mimics human visual system, good for texture analysis

6. GIST Features

• Purpose: Global scene descriptors
• Feature Dimension: 512 features
• Advantage: Captures global scene properties and spatial layout

🤖 Clustering Approaches

K-Means Clustering

• Usage: Initial clustering experiments and baseline comparison
• Parameters: n_clusters=2-5, random_state=42
• Evaluation: Silhouette scores, cluster visualization
• Limitations: Requires predefined number of clusters, sensitive to initialization

DBSCAN Clustering (Primary Method)

• Usage: Advanced clustering with automatic noise detection
• Parameter Tuning: Systematic eps optimization (0.1-2.0 range)
• Advantages:
  • No need to specify number of clusters
  • Handles noise and outliers effectively
  • Discovers clusters of arbitrary shapes
• Evaluation: Silhouette scores, NMI, Rand Index

📈 Performance Evaluation Metrics

1. Silhouette Score

• Purpose: Measures cluster cohesion and separation
• Range: -1 to 1 (higher is better)
• Interpretation: Measures how similar an object is to its own cluster vs other clusters

2. Normalized Mutual Information (NMI)

• Purpose: Measures cluster-label agreement
• Range: 0 to 1 (higher is better)
• Interpretation: Quantifies the mutual information between true and predicted clusters

3. Rand Index (RI)

• Purpose: Overall clustering accuracy
• Range: 0 to 1 (higher is better)
• Interpretation: Measures the percentage of correct decisions made by the algorithm

🔬 Research Methodology

Phase 1: Data Preparation

1. Image Loading: Load images from directory structure
2. Preprocessing: Convert to grayscale, normalize pixel values
3. Label Extraction: Extract true labels from directory structure
4. Data Organization: Organize files for systematic processing

Phase 2: Feature Extraction

1. Individual Feature Extraction: Extract each feature type separately
2. Feature Standardization: Apply StandardScaler for normalization
3. Dimensionality Reduction: Apply PCA to reduce to 2D for visualization
4. Feature Combination: Create various combinations of features

Phase 3: Clustering Experiments

1. Parameter Tuning: Systematic optimization of DBSCAN parameters
2. Multiple Runs: Test different feature combinations
3. Performance Evaluation: Compute multiple evaluation metrics
4. Statistical Analysis: Compare results across different configurations

Phase 4: Analysis & Visualization

1. 2D Visualization: PCA plots showing cluster distributions
2. Performance Comparison: Tabular comparison of results
3. Statistical Significance: Analysis of performance differences
4. Result Interpretation: Draw conclusions from experimental data

📊 Key Research Findings

Best Performing Configuration

• Feature Combination: LBP + WM + CENTRIST (60 features total)
• Clustering Method: DBSCAN (eps=1.1, min_samples=20)
• Performance Metrics:
  • NMI: 0.320
  • Rand Index: 0.720
  • Silhouette Score: 0.454
• Clusters Discovered: 2 main clusters with 245 outliers

Performance Comparison Table

Feature Combination	Best EPS	Clusters	NMI	Rand Index	Silhouette Score
LBP only	0.7	2	0.299	0.708	0.412
WM only	0.2	3	0.282	0.665	0.096
CENTRIST only	0.7	2	0.299	0.708	0.412
LBP + WM	0.6	2	0.289	0.673	0.305
WM + CENTRIST	0.6	2	0.289	0.673	0.305
LBP + CENTRIST	1.0	2	0.302	0.710	0.415
LBP + WM + CENTRIST	1.1	2	0.320	0.720	0.454

Key Insights

1. Feature Complementarity:

   • LBP and CENTRIST show similar individual performance
   • WM features provide complementary information when combined
   • Three-feature combination achieves optimal performance

2. Clustering Algorithm Performance:

   • DBSCAN outperforms K-Means due to noise handling
   • Parameter sensitivity is critical for DBSCAN performance
   • eps parameter optimization is essential for good results

3. Dataset Characteristics:

   • 2-3 clusters optimal for Caltech-7 dataset
   • Significant number of outliers present (245 out of 1474)
   • Feature standardization crucial for multi-view clustering

🗂️ Repository Structure

caltech/
├── README.md                           # Project documentation
├── requirements.txt                    # Python dependencies
├── .gitignore                         # Git ignore file
├── CALTECH7_DB-SCAN.pdf              # Research results and analysis
│
├── notebooks/                         # Jupyter notebooks
│   ├── CALTECH7LBP-checkpoint.ipynb  # LBP analysis on Caltech-7
│   ├── finalcal7-checkpoint.ipynb    # Comprehensive Caltech-7 analysis
│   ├── DBSCAN-checkpoint.ipynb       # DBSCAN parameter optimization
│   ├── car_coil20-checkpoint.ipynb   # COIL-20 car subset analysis
│   ├── COIL20_Image_Processing-checkpoint.ipynb  # Full COIL-20 processing
│   ├── combined_class_coik_correct-checkpoint.ipynb  # Combined features
│   ├── dbscan_car-checkpoint.ipynb   # DBSCAN on car subset
│   ├── hog_class_coil_correct-checkpoint.ipynb  # HOG feature analysis
│   ├── lbp_class_coil_correct-checkpoint.ipynb  # LBP feature analysis
│   ├── fileCALTECH7-checkpoint.ipynb # Dataset organization
│   └── filesharing-checkpoint.ipynb  # File management utilities
│
├── data/                              # Dataset information
│   ├── caltech7_info.md              # Caltech-7 dataset details
│   └── coil20_info.md                # COIL-20 dataset details
│
├── results/                           # Results and analysis
│   └── performance_summary.csv       # Performance metrics table
│
└── src/                              # Source code modules (empty)

🚀 Getting Started

Prerequisites

• Python 3.7+
• Jupyter Notebook
• Required Python packages (see requirements.txt)

Installation

1. Clone the repository:

git clone https://github.com/yourusername/caltech-multiview-clustering.git
cd caltech-multiview-clustering

2. Install dependencies:

pip install -r requirements.txt

3. Download datasets:
   • Caltech-7: Place in data/caltech7/
   • COIL-20: Place in data/coil20/

Usage

1. Run individual experiments:

   • Open any notebook in the notebooks/ directory
   • Follow the execution order in each notebook

2. Reproduce results:

   • Start with finalcal7-checkpoint.ipynb for comprehensive analysis
   • Use DBSCAN-checkpoint.ipynb for parameter optimization

3. Custom experiments:

   • Modify feature extraction parameters in src/feature_extraction.py
   • Adjust clustering parameters in src/clustering.py

📝 Notebook Descriptions

Core Analysis Notebooks

• finalcal7-checkpoint.ipynb: Main research notebook with comprehensive analysis
• CALTECH7LBP-checkpoint.ipynb: LBP feature analysis on Caltech-7
• DBSCAN-checkpoint.ipynb: DBSCAN parameter optimization and analysis

Feature-Specific Notebooks

• hog_class_coil_correct-checkpoint.ipynb: HOG feature extraction and clustering
• lbp_class_coil_correct-checkpoint.ipynb: LBP feature extraction and clustering
• combined_class_coik_correct-checkpoint.ipynb: Combined feature analysis

Dataset-Specific Notebooks

• car_coil20-checkpoint.ipynb: COIL-20 car subset analysis
• COIL20_Image_Processing-checkpoint.ipynb: Full COIL-20 dataset processing

Utility Notebooks

• fileCALTECH7-checkpoint.ipynb: Dataset organization and preparation
• filesharing-checkpoint.ipynb: File management and organization

🔬 Technical Implementation Details

Feature Extraction Pipeline

1. Image Preprocessing: Grayscale conversion, normalization
2. Feature Computation: Parallel extraction of multiple feature types
3. Feature Standardization: Z-score normalization
4. Dimensionality Reduction: PCA for visualization (2D)

Clustering Pipeline

1. Parameter Optimization: Grid search for optimal eps values
2. Clustering Execution: DBSCAN with optimized parameters
3. Performance Evaluation: Multiple metrics computation
4. Visualization: 2D PCA plots with cluster assignments

Evaluation Pipeline

1. Metric Computation: Silhouette score, NMI, Rand Index
2. Statistical Analysis: Performance comparison across configurations
3. Visualization: Cluster plots, performance charts
4. Interpretation: Results analysis and insights

📚 Research Context and Motivation

Why Multi-View Clustering?

Multi-view clustering leverages multiple representations (views) of the same data to improve clustering performance. Each feature extraction method captures different aspects of the data:

• LBP: Local texture patterns
• HOG: Shape and edge information
• WM: Multi-resolution characteristics
• CENTRIST: Census transform features

Why DBSCAN?

DBSCAN was chosen over K-Means because:

1. No need to specify cluster count: Automatically discovers optimal number of clusters
2. Noise handling: Identifies and handles outliers effectively
3. Shape flexibility: Can find clusters of arbitrary shapes
4. Parameter robustness: Less sensitive to parameter variations

Why These Datasets?

• Caltech-7: Provides diverse object categories with varying complexity
• COIL-20: Offers controlled rotation variations for validation
• Combined analysis: Allows for comprehensive evaluation across different data characteristics

🎓 Research Contributions

1. Systematic Evaluation: Comprehensive comparison of 6 different feature extraction methods
2. Multi-View Analysis: Investigation of feature combination effectiveness
3. Parameter Optimization: Systematic DBSCAN parameter tuning methodology
4. Performance Benchmarking: Detailed evaluation using multiple metrics
5. Reproducible Research: Complete code and data organization for reproducibility

🔮 Future Work and Extensions

Potential Improvements

1. Additional Feature Types: Include more advanced features (CNN features, SIFT, etc.)
2. Advanced Clustering: Test other clustering algorithms (Spectral Clustering, Gaussian Mixture Models)
3. Multi-View Fusion: Implement advanced multi-view fusion techniques
4. Larger Datasets: Extend analysis to larger image datasets
5. Deep Learning Integration: Incorporate deep learning-based feature extraction

Research Directions

1. Theoretical Analysis: Mathematical analysis of multi-view clustering performance
2. Scalability: Optimization for large-scale datasets
3. Real-world Applications: Apply to practical computer vision problems
4. Comparative Studies: Compare with state-of-the-art multi-view clustering methods

📞 Contact

For questions or collaboration opportunities:

• Email: kunalsali04@gmail.com
• GitHub: kunnaall04

---

Note: This research demonstrates multi-view clustering using various feature extraction methods and DBSCAN clustering on image datasets.