Skip to content

Industrial color difference evaluation using a Neural Network and a Fuzzy Inference System

Notifications You must be signed in to change notification settings

psoftware/IS-color-diff-nn

Repository files navigation

Color Diff Neural Network/FIS - Intelligent System Course Project

The goal of this project is to build a system that can tell "how much" a color is different from a sample color. This is useful for industrial companies that want to check print color accuracy, considering how this differences are perceived by customers.

More details

However, colors are not perceived in the same way for everyone. To solve this we basically used CIE Standards and Lab/LCh color space. Our first solution was to implement a Fitting Neural Network and use a simple Euclidean Distance to train the network. However this turned to be wrong because Euclidean Distance does not work well when facing some parts of the LCh color space, because it tends to overstimate color differences. To solve this we added a Fuzzy Inference System that does a correction on DE when needed (for example when color is dark, when color is almost purple, and other cases).

Master/Copy generation

To train the network we need a couple of master and copy colors: master ones are provided as project material, but copies must be generated by us. How we generate copies? Our choice is to take the master color spectrum, divide it in some parts, and sum a value for each part, for each element of that part. This choice is justified by how spectrum to perceived color function works. Other details can be found into the main project doc.

Results

10 features (so 10 Input Layer Neurons), 1 Hidden Layer of 8 neurons

R ~= 0.99, mse ~= 0.02

Check the main doc for more detailed results.

About

Industrial color difference evaluation using a Neural Network and a Fuzzy Inference System

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages