Scratch Removal Tool

Before	After

Overview

This Scratch Removal Tool is the first part of a larger image restoration project. It provides intelligent detection and removal of scratches from images using advanced computer vision techniques. The application combines deep learning-based scratch detection with inpainting algorithms to produce high-quality restored images.

Environment

This application uses a Python environment similar to those used for DeOldify or ESRGAN projects. The primary dependencies include:

Python 3.10+
PyTorch 1.10
OpenCV
FastAPI
Streamlit
PIL (Pillow)

Key Features

Scratch Detection: Automatically identifies scratches in images using a deep learning model
Scratch Removal: Applies state-of-the-art inpainting algorithms to remove detected scratches
Multiple Inpainting Methods: Choose between TELEA, Navier-Stokes (NS), or LaMa (AI-based)
Adjustable Parameters: Fine-tune threshold, radius, and dilation for optimal results
Web UI: User-friendly interface for uploading and processing images

Project Structure

api.py - FastAPI implementation for the backend services
app.py - Streamlit web application for the user interface
detection.py - Deep learning model for scratch detection
removal.py - Implementation of scratch removal algorithms

Getting Started

Installation

Clone the repository
Create a compatible environment ( preferred with conda )

Install dependencies:

Using conda:

conda env create -f environment.yml
conda activate venv

Running the Application

To start both the API server and web interface:

python api.py

And in another terminal:

python -m streamlit run app.py

API Usage

The API provides several endpoints:

/detect/ - Detect scratches in an image
/remove/ - Remove scratches using a provided mask
/detect-and-remove/ - Perform both detection and removal in one step
/status/ - Check API server status and resource usage

Example API call:

import requests

# Detect and remove scratches
url = "http://localhost:8000/detect-and-remove/"
files = {"file": open("image.jpg", "rb")}
params = {
    "method": "telea",
    "radius": 2,
    "threshold": 47,
    "dilation": 1,
    "use_dilation": True,
    "use_lama": False  # Set to True to use LAMA AI-based inpainting
}

response = requests.post(url, files=files, data=params)
with open("restored.png", "wb") as f:
    f.write(response.content)

Web Interface Usage

Open the web interface at http://localhost:8501
Upload an image with scratches
Click "Detect Scratches" to visualize the detected scratches
Adjust parameters if needed (including selecting LAMA for higher quality)
Click "Remove Scratches" to perform restoration
Download the restored image

Customization Options

Inpainting Method:
- telea - Alexandru Telea algorithm (better for thin scratches)
- ns - Navier-Stokes based method (better for larger areas)
- LAMA - AI-based inpainting (best quality but slower)
Radius: Pixel radius for the inpainting algorithm (1-10, only for OpenCV methods)
Mask Threshold: Sensitivity of scratch detection (1-254)
Dilation: Expand detected scratch areas for better coverage (1-5)

Technical Details

Scratch Detection

The detection system uses a U-Net architecture trained on a dataset of scratched images. It identifies scratches with high precision, producing a binary mask that highlights damaged areas.

NOTE: the detection pretrained model retrieved from microsoft bring old photos back to life project releases.

The detection algorithm (detection.py) works by:

Loading a pre-trained U-Net model from the checkpoints/detection/ directory
Processing input images to the appropriate size
Running inference to produce probability maps highlighting scratch regions
Post-processing these maps to create binary masks
Saving the resulting masks as PNG files

Scratch Removal

The removal process uses inpainting algorithms to reconstruct the damaged areas based on surrounding pixels. The tool supports:

TELEA Algorithm: Fast and effective for thin scratches
Navier-Stokes: Better quality for larger damaged areas, but slower
LAMA: AI-based inpainting that provides the highest quality results, especially for complex textures

LAMA Integration

LAMA (Large Mask Inpainting) is an advanced AI-based inpainting algorithm that has been integrated into the tool. It often provides superior results compared to traditional methods, especially for larger areas and complex textures.

How it works: LAMA uses a deep learning model to fill in masked regions with content that is contextually appropriate
Performance: LAMA processing is slower than OpenCV methods but produces higher quality results
Fallback: If LAMA fails for any reason, the system will automatically fall back to the selected OpenCV method

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Acknowledgments

OpenCV for image processing capabilities
PyTorch for the deep learning framework
FastAPI and Streamlit for the web components
LaMa for the advanced inpainting algorithm
Microsoft's "Bringing Old Photos Back to Life" project for the detection model