Image Colorization with Neural Networks

A deep learning application that colorizes grayscale images, deployed on Kubernetes with a full CI/CD pipeline.

---

Live Demo
https://colorization.dyn.cloud.e-infra.cz/

📖 Table of Contents

• Project Overview
• Key Features
• Architecture
• Tech Stack
• The Machine Learning Model
• Project Repositories

🌟 Project Overview

I wanted this project to be complete, end-to-end process which goes beyond just training a deep learning model. The goal is to bring the model to user-facing frontend in a scalable and devops-focused way. It takes a grayscale image from a user, uses a Conditional Generative Adversarial Network (cGAN) to add realistic color, and returns the result. The entire system is containerized and deployed on a public Kubernetes cluster using Helm, with automated builds and deployments managed by GitHub Actions.

✨ Key Features

• ML-Powered Colorization: Utilizes a Pix2Pix cGAN model trained on the COCO dataset.
• RESTful API: A simple and robust FastAPI backend provides the inference endpoint.
• Interactive Frontend: A SvelteKit frontend for interacting with the model.
• Scalable Infrastructure: Deployed on Kubernetes for high availability and scalability.
• Automated CI/CD: GitHub Actions pipeline automatically builds, and deploys changes to the cluster.

🏗️ Architecture

The system is designed as a set of decoupled microservices that communicate over HTTP and a message queue for robust, asynchronous processing.

1. User uploads an image via the SvelteKit Frontend.
2. The Frontend sends the image to the FastAPI Backend API endpoint.
3. The Backend does the inference using the provided grayscale image.
4. The resulting colorized image is sent back to the frontend.

🛠️ Tech Stack

Category	Technology
Frontend	SvelteKit, TypeScript, TailwindCSS
Backend	Python, FastAPI, Pydantic
ML/Data Science	PyTorch, NumPy, Pillow
Infrastructure	Docker, Kubernetes, Helm, NGINX Ingress
CI/CD	GitHub Actions

🧠 The Machine Learning Model

I was following the protocol proposed in the Pix2Pix paper for training the Conditional GAN model. The architecture consists of a U-Net generator and a PatchGAN discriminator.

• Generator: U-Net architecture with skip connections tasked with generating detailed images.
• Discriminator: PatchGAN classifier that determines if (input, output) pairs are real or fake.
• Dataset: Trained on a 10,000-image subset of the COCO dataset.
• https://github.com/vojtam/cGAN-image-colorization

📂 Project Repositories

This project is split into four repositories, each with a specific responsibility:

Repository	Description	Link
🎨 Frontend	SvelteKit user interface.	image-colorizer-frontend
🚀 Backend	FastAPI application handling API requests.	(This Repository)
🧠 Model Training	Scripts, notebooks, and code for training the cGAN model.	image-colorizer-training
⚙️ Infrastructure	Kubernetes manifests, Helm charts, and CI/CD workflows.	image-colorizer-infra