evoimage

evoimage

Simple attempt at evolving image reproductions

Overview

This project is experimental and for fun. It doesn't really accomplish anything that is publication worthy, but you might enjoy playing with it. See main.go for command line options and the main loop. See representation.go for the main representation and encoding.

Implementation

The main loop is a standard generational GA, but includes some adaptations during the run. (Note that we equate "making progress" with an increase in the best individual - which is a decrease is our minimized fitness score).

• We use elitism (see Elitism below)
• Mutation rate is increased when progress has stalled (see Mutation below)
• The population size increases every generation progress is not made, but returns to the default level when progress resumes
• Tournament size is rotated (see Selection below)

Fitness Function

The fitness function is the sum of the Euclidean distance in RGB space for all pixels. Note that this means we are attempting to minimize our fitness function.

See representation.go.

Representation

Each individual is an ordered list of genes where each gene is a triangle and a color in RGBA space (note our use of transparency for representation and drawing, but not in the final fitness function).

Color and spatial coordinates are sampled uniformly at random when creating a random instance.

See representation.go.

Selection

Selection is currently tournament selection. The main loop uses a rotating tournament size (2-5 inclusive).

See selection.go and main.go.

Mutation

We use Gaussian mutation. Currently the standard deviation is different for spatial and color coordinates and is hand-coded.

The mutation rate currently receives a (capped) increase for every generation since we have failed to get an increase in the best fitness score.

There is also a gene shuffle operator used as part of our elitism strategy (see below).

See mutation.go.

Crossover

Crossover is uniform crossover.

See crossover.go.

Elitism

We copy the five best individuals to the next generation.

We also add a second copy of the best five individuals, but with their genes shuffled and mutated.

Installing

You can just use the executable in the folder matching your operating system and architecture under dist.

Building and running

To build this project, you need Go 1.9 and recent version of GNU Make running in a bash shell (modern bash is preferred). All builds and runs have been performed on Ubuntu 16.04 and 17.04, so everything you need can be installed via apt. However, please note that you will need to install Go 1.9 (or later) manually. You can use ubuntu-make for this - see https://github.com/ubuntu/ubuntu-make

Go is cross-platform, so you should be able to build on Windows or Mac OSX, bu that has not been tested.

Go dependencies are managed with dep (see https://github.com/golang/dep for details), but you shouldn't need to worry about that unless you are adding or upgrading dependencies.

See Makefile. The short version is that static source code analysis and formatting can be checked with make lint. Build with make.

After building, run ./evoimage -h to see all parameter options.

To run on an image with all default parameters, you only need to supply the -image parameter. For example, ./evoimage -image imgs/target-mondrian.jpg.

evoimage appends to log files (named for the target image) in the log directory during the run. It also writes the best image for each generation to the output directory. Image name formats use the generation number but not the file name, so you should save or clear the output directory before starting a new run. We also write the best image as ./latest.jpg.

Running ./script/output_ani will take all current images in the output directory and create an mp4 video showing progress. Note that ffmpeg must be installed.

Images

• "Portrait of Isabel Parreño y Arce, Marquesa de Llano, Anton Raphael Mengs, 1771 - 1772" by Anton Raphael Mengs is licensed under CC0 1.0 https://ccsearch.creativecommons.org/image/detail/TJmttlM53HMeELT3IduwGw==
• "Landscape" by Carle (Antoine Charles Horace) Vernet (French, Bordeaux 1758–1836 Paris) via The Metropolitan Museum of Art is licensed under CC0 1.0 https://ccsearch.creativecommons.org/image/detail/7J6dR75Zr50Kwq7_Ax0ehw==
• "Composition" by Piet Mondrian (Dutch, Amersfoort 1872–1944 New York) via The Metropolitan Museum of Art is licensed under CC0 1.0 https://ccsearch.creativecommons.org/image/detail/wGM3NBEVzE0rIfI48n1NIQ==
• "mondrian" by apenny is licensed under CC BY 2.0 https://ccsearch.creativecommons.org/image/detail/dci6KvaAj9FaCqASttXsEQ==
• "Versailles" by Auguste Renoir (French, Limoges 1841–1919 Cagnes-sur-Mer) via The Metropolitan Museum of Art is licensed under CC0 1.0 https://ccsearch.creativecommons.org/image/detail/LZKw8dOJS85mFqXEMeWrwA==
• "The Milliner" by Auguste Renoir (French, Limoges 1841–1919 Cagnes-sur-Mer) via The Metropolitan Museum of Art is licensed under CC0 1.0 https://ccsearch.creativecommons.org/image/detail/6DhumcJp7g8tXzWCgIfjog==

The original full size CC JPEG's from above are in the imgs directory with the prefix orig-. They have been scaled down so the largest dimension is 256 pixels while preserving the original aspect ratio. The scaled images used as target images are also in the imgs directory with the prefix target.