Color vision deficiency datasets & recoloring evaluation using GANs

People with Color Vision Deficiency (CVD) cannot distinguish some color combinations under normal situations. Recoloring becomes a necessary adaptation procedure. In this paper, in order to adaptively find the key color components in an image, we first propose a self-adapting recoloring method with an Improved Octree Quantification Method (IOQM). Second, we design a screening tool of CVD datasets that is used to integrate multiple recoloring methods. Third, a CVD dataset is constructed with the help of our designed screening tool. Our dataset consists of 2313 pairs of training images and 771 pairs of testing images. Fourth, multiple GANs i.e., pix2pix-GAN [1], Cycle-GAN [2], Bicycle-GAN [3] are used for colorblind data conversion. This is the first ever effort in this research area using GANs. Experimental results show that pix2pix-GAN [1] can effectively recolor unrecognizable colors for people with CVD, and we predict that this dataset can provide some help for color blind images recoloring. Datasets and source are available at: https://github.com/doubletry/pix2pix, https://github.com/doubletry/CycleGAN and https://github.com/doubletry/BicycleGAN.