Linear-Colour-Transfer

Simple, effective colour transfer using linear algebra. Supports arbitrary image masking for selective transfer.

• Given a source image (reference) and a target image (image to be recolored) this repo performs fast linear colour transfer based on similarity transform in RGB vector space.
• based upon the histogram matching method outlined in Gatys, Leon A., et al., "Controlling perceptual factors in neural style transfer." IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017."
• Code is adapted from and extends the code in the repo of https://github.com/ProGamerGov/Neural-Tools.git.
• The primary extension is that now colour can be arbitrarily sampled from a reference image using binary masking and applied selectively to a target image also using binary masking (see examples).
• This method is much more stable than simple matching of mean and standard deviation (see baseline example comparison).

Examples

Baseline Example

Example 1: Full colour transfer based on a reference image

Example 2: Selective recoloring using image masking on the target image

Example 3: Selective color sampling from reference image and selective recoloring on the target image