README.md

Prokudin-Gorskii Image Alignment

This project implements automatic alignment of Prokudin-Gorskii glass plate photographs using computer vision techniques.

Overview

The Prokudin-Gorskii collection contains early color photographs taken by Sergei Mikhailovich Prokudin-Gorskii between 1907-1915. Each image consists of three separate exposures (Blue, Green, Red channels) that need to be aligned to create a single color photograph.

Features

• Single-scale alignment for small JPEG images using exhaustive search
• Multi-scale pyramid alignment for large TIFF images using coarse-to-fine optimization
• Edge-based features using Sobel gradients for robust alignment
• Normalized Cross-Correlation (NCC) scoring metric
• Bells & Whistles enhancements:
  • Automatic border cropping
  • White balance correction
  • Automatic contrast adjustment

Requirements

pip install Pillow scipy numpy

Usage

Basic Usage

from main import ProkudinGorskiiAligner

# Create aligner instance
aligner = ProkudinGorskiiAligner()

# Process a single image
aligner.process("images/cathedral.jpg", "output/cathedral_restored.jpg")

Batch Processing

# Process all images in a directory
aligner.batch_process("images", "output")

# Process specific images
subset = ["cathedral.jpg", "monastery.jpg", "emir.tif"]
aligner.batch_process("images", "output", file_list=subset)

Command Line

python main.py

Algorithm Details

Single-Scale Alignment (JPEGs)

• Used for small images (cathedral.jpg, monastery.jpg, tobolsk.jpg)
• Exhaustive search over displacement window [-15, 15] pixels
• Direct application of NCC metric on Sobel edge features

Multi-Scale Pyramid Alignment (TIFFs)

• Used for large images (emir.tif, harvesters.tif, etc.)
• Recursive coarse-to-fine alignment across 5 pyramid levels
• 2x downsampling at each level using scipy.ndimage.zoom
• Refines alignment with ±2 pixel search at each level

Edge Features

• Sobel gradient magnitude: sqrt(dx² + dy²) using scipy.ndimage.sobel
• Robust to brightness differences between channels
• Essential for problematic images like emir.tif

Implementation

• Image I/O: PIL (Pillow) for reading and writing images
• Edge detection: scipy.ndimage.sobel for gradient computation
• Image resizing: scipy.ndimage.zoom for pyramid downsampling
• Normalization: numpy operations for 8-bit conversion

File Structure

├── main.py              # Main implementation
├── images/               # Input images directory
│   ├── cathedral.jpg
│   ├── monastery.jpg
│   ├── emir.tif
│   └── ...
├── output/               # Generated output directory
├── index.html           # Results webpage
└── README.md           # This file

Parameters

• search_range=15: Maximum displacement search window
• pyramid_depth=5: Number of pyramid levels for multi-scale alignment
• crop_ratio=0.15: Border cropping ratio to avoid edge artifacts

Results

See index.html for visual results and computed alignment offsets for all processed images.

Author

Mohak Sharma (ms7306)