🧬 Skin Lesion Classification Pipeline (HAM10000)

This project performs feature-based classification of skin lesions using the HAM10000 dataset. It combines classic computer vision techniques (HOG and LBP), dimensionality reduction (PCA), and traditional machine learning models (SVM, Softmax Regression, OvR) to classify dermatological images.

---

🔍 Project Overview

• 📊 Dataset: HAM10000
• 🧠 Features:
  • HOG (Histogram of Oriented Gradients)
  • LBP (Local Binary Patterns)
  • PCA for dimensionality reduction
• 🧪 Models:
  • SVM (with and without PCA)
  • Logistic Regression (Softmax)
  • Logistic Regression (OvR)
• 📉 Evaluation:
  • Confusion Matrix
  • Classification Report (Precision, Recall, F1-score)

---

🧰 Tools & Libraries

• Python
• OpenCV
• scikit-image
• scikit-learn
• TensorFlow
• CuML (GPU-accelerated SVM)
• Matplotlib, Seaborn

---

⚙️ Feature Extraction Pipeline

1. Resize images to 64x64
2. Convert to grayscale
3. Extract:
   • HOG features (orientation, blocks, cells)
   • LBP histograms (radius, points, uniform)
4. Combine features
5. Apply PCA to reduce to 50 components

---

📈 Classifier Comparison

Classifier	Input	Accuracy	Notes
SVM	Raw Images	Low	High dimensionality, weak performance
SVM	PCA	Best	High accuracy, robust classification
Logistic (Softmax)	Raw/PCA	Medium	Sensitive to feature quality
Logistic (OvR)	Raw	Very Low	Weak prediction across all classes
Logistic (OvR)	PCA	Medium	PCA improves OvR significantly

---

🖼️ Visual Outputs

• Confusion matrices for each model
• HOG image visualizations
• PCA explained variance plot

---

🧪 Results Summary

• SVM with PCA reached up to 90% accuracy
• PCA significantly boosts performance across models
• Feature extraction + PCA is crucial for good ML performance on image data

---