Image Colorization using Autoencoder

This project implements an image colorization model using a deep learning approach with an Autoencoder architecture. The model takes grayscale images as input and predicts the missing color channels.

Table of Contents

• Project Overview
• Dataset
• Model Architecture
• Training Strategy
• Results
• Usage
• Requirements
• How to Run
• References

---

Project Overview

The goal of this project is to train a neural network that can take grayscale images (L channel from Lab color space) and predict the missing a and b color channels. The model is trained using Mean Squared Error (MSE) loss or Structural Similarity Index (SSIM) loss.

Dataset

The dataset consists of images converted to the Lab color space, where:

• L channel represents lightness (input to the model)
• a, b channels represent color information (target output)

The dataset is split into:

• 70% Training set
• 15% Validation set
• 15% Test set

Model Architecture

The model is based on an Autoencoder with a U-Net-like structure:

• Encoder: Extracts features from the grayscale input
• Decoder: Reconstructs the a and b channels
• Activation Function: tanh for output normalization

Hyperparameters

• Optimizer: Adam
• Loss Function: SSIM Loss
• Learning Rate: 0.001
• Epochs: 100
• Batch Size: Tuned using cross-validation

Training Strategy

• Cross-validation is used to find the best hyperparameters
• Early stopping monitors SSIM loss to prevent overfitting
• Metrics:
  • Peak Signal-to-Noise Ratio (PSNR)
  • Structural Similarity Index (SSIM)

Training and Validation Metrics

After training, the following plots are generated:

1. Loss curves (Training & Validation)
2. SSIM metric over epochs
3. PSNR metric over epochs

Results

The final model is evaluated by predicting color channels and reconstructing RGB images. The outputs include:

• Original image (ground truth)
• Grayscale input (L channel)
• Predicted a and b channels
• Reconstructed RGB image

Usage

Installation

pip install -r requirements.txt

Requirements

• Python 3.x
• TensorFlow / Keras
• OpenCV
• NumPy
• Matplotlib

How to Run

1. Clone this repository:

   git clone https://github.com/yourusername/image-colorization.git
   cd image-colorization

2. Install dependencies:

   pip install -r requirements.txt

References

• Kaggle: Image Colorization with CNN
• Research Papers on Autoencoders for Image Processing

---