Automatic Processing of DSLR-Scanned Film Negatives

This repository contains the code surrounding the Computational Photography course project aiming to automate the processingof DSLR-scanned film negatives.

Project Structure

.
├── data/                   # Data storage and organization
│   ├── raw/               # Raw camera files (.RAF, .HIF)
│   ├── prepro/            # Pre-processed images  
│   ├── datasets/          # Training/validation datasets
│   ├── checkpoints/       # Model checkpoints
│   └── reproduction/      # Reproduction assets
├── src/                   # Source code
│   ├── nn/               # Machine Learning approach
│   │   ├── models/       # Neural network architectures
│   │   ├── utils/        # Training utilities and dataset loading
│   │   ├── eval/         # Evaluation and testing tools
│   │   └── data/         # Data handling
│   ├── scripts/          # Processing and utility scripts. Includes training scripts.
│   │   ├── sigmoid/      # Statistical sigmoid correction
│   │   ├── kp_calc/      # Tool to collect the color dataset by
│   │   └── train/        # Training scripts
│   ├── processing/       # Statistical image processing. Contains the Sigmoid correct code
│   ├── analysis/         # Analysis and visualization tools
│   ├── utils/           # General utilities
└── requirements.txt

Machine Learning Approach (src/nn/)

The neural network pipeline provides end-to-end learning for film negative correction:

• The model architectures are located in src/nn/models (one_hidden, enhanced_comb, combhidden)
• Full training pipeline with validation and testing through the main_trainer.py script (see src/scripts/training for an example invocation)
  • Training in both RGB and LAB color spaces has been tested (can be switched with a flag)

Statistical Processing (src/scripts/sigmoid/ and src/processing/)

• White Balance methods can be selected using the WhiteBalance class (white_balance.py).
• Hyper-parameter optimization (namely generating a PDF with various combinations of hyperparameters, which can be selected through a configuration file) through the src/scripts/sigmoid/correct_many.py

Data Collection Tool (src/scripts/kp_calc/)

Interactive GUI application for data collection and annotation

• Using image mappings, draw matching rectangles to select pixel regions across image sets
• Adapts based on image transforms
• Used to generate CSV datasets for training the NN models (see --multiple and --data-processing flags)

Usage:

python src/scripts/kp_calc/app.py

Recommended setup

Prerequisites

• Python 3.9+ (recommended 3.12+, tested with 3.10+)
• PyTorch if NN training(install according to your system: https://pytorch.org/get-started/locally/)
• exiftool (required for RAW file metadata extraction)

Installation

1. Clone the repository:

git clone https://github.com/vigarov/positron
cd positron

2. Install dependencies:

pip install -r requirements.txt

3. Set up data directory structure (see data/README.md for details)

Quick Start Workflow

1. Data Collection: Use the interactive tool to collect training data

python src/scripts/kp_calc/kp_calc.py -h

2. Statistical Processing: Generate a PDF with applying the sigmoid correction method on all the images from the dataset

python src/scripts/sigmoid/correct_many.py -h

3. ML Training: Train a neural network model

Simply execute one of the trainers in src/scripts/train/ or run your own train by invoking the train loop:

python src/nn/main_trainer.py -h