Genetic Image Evolution

A Kotlin implementation of evolutionary image generation using genetic algorithms¹ to evolve geometric shapes that approximate a target image.

Overview

This project uses genetic programming to evolve simple geometric shapes (circles, rectangles, polygons, curves) that gradually approximate a target image through natural selection. Each "individual" consists of a genome of coloured shapes, and evolution occurs through mutation and fitness-based selection.

Inspiration

This implementation is inspired by:

• Roger Johansson's "Genetic Programming: Evolution of Mona Lisa"
• NP Contemplation's "Clojure genetic Mona Lisa problem"

How It Works

1. Initialization: Start with a random collection of zero or more geometric shapes
2. Mutation: Randomly modify shape properties (position, colour, size) or add new shapes
3. Fitness Evaluation: Compare the rendered image against the target using pixel-by-pixel RGB distance
4. Selection: Keep mutations that improve fitness (lower pixel distance)
5. Iteration: Repeat the process continuously to evolve better approximations

Configuration

Evolution parameters can be configured via evolver-settings.json:

• maxGenomeSize - Maximum number of shapes per individual
• newShapeProbabilityFactor - Likelihood of adding new shapes
• avgShapesToMutate - Average number of shapes to mutate per generation

Architecture

The system supports both local evolution; a GUI will show you the latest state if a DISPLAY exists but not if run headless. Shapes can be automatically uploaded to the (appallingly named) dudestore web service.

---

Footnotes

¹ Note that I was later informed that this isn't a true genetic algorithm and is, in fact, a stochastic hill climbing algorithm. I confess I'm no expert on either and continue to call this project "genetic", even if it's inaccurate.