Training a neural network to approximate a novel color gradient function based on CIEDE2000.
CIEDE2000 - a color difference formula based on human vision and used across industry.
Color gradient - an interpolation between two colors; a "path" from one color to another.
A novel gradient function has been developed based on CIEDE2000. It sometimes produces non-continuous gradients. Approximating it computationally requires a relatively long time.
This function was used to train a neural network, successfully reducing computation time while capturing its intricacies like non-continuity.
This repository contains the source code for the tools that were used to achieve this result:
- sample-generator - generates training and testing data for the neural network
- nn - creates, trains and tests a neural network
- image-generator - a tool for visualizing how the neural network compares to the original function
Requirements:
- Rust toolchain
- SDL2 developer libraries (for image-generator only)
cd sample-generator
cargo run --releasecd nn
cargo run --releasecd image-generator
cargo run --releaseNote that sample-generator takes about 10 minutes to complete while using the default configuration.
sample-generator produces data-[seed].csv and nn produces nn.json.
All 3 tools have config.ini files that can be edited to change their configuration.
For convenience, nn uses provided sample_data.csv and image-generator uses provided sample_nn.json. This can be changed to user provided data.cvs and nn.json by setting use_sample_data and use_sample_nn to false in the respective configuration files.
Top rows - a slow computational approximation of the CIEDE2000 gradient between 2 random colors. Bottom rows - a fast neural network approximation of the same.
The neural network used for the generation of these images is contained in image-generator/sample_nn.json. It was trained with 10.9 million samples for 80 epochs.
