Color Universal Design Neural Network for the Color Vision Deficiencies

TL;DR

This work tackles making images readable to color-deficient viewers by introducing CUD-Net, a per-image neural-filter that learns a small, piecewise-linear filter to simultaneously preserve colors and increase contrast for non-CUD objects. It advances data preparation (two refined groups and deuteranopia augmentation), a multi-modality fusion backbone using three inputs ($I^n$, $I^d$, $I^m$) and a per-image regulator with $M=64$ node points, and a conjugate loss framework to handle one-to-many mappings while maintaining color fidelity. Empirical results show CUD-Net outperforms GAN-based and neural-filter baselines in deuteranopia and protanopia scenarios, with a user study indicating higher object distinguishability and favorable color harmony. The approach holds practical value for high-resolution, publication-ready CUD images and enables future web/API deployment for real-time color-deficiency accommodation.

Abstract

Information regarding images should be visually understood by anyone, including those with color deficiency. However, such information is not recognizable if the color that seems to be distorted to the color deficiencies meets an adjacent object. The aim of this paper is to propose a color universal design network, called CUD-Net, that generates images that are visually understandable by individuals with color deficiency. CUD-Net is a convolutional deep neural network that can preserve color and distinguish colors for input images by regressing the node point of a piecewise linear function and using a specific filter for each image. To generate CUD images for color deficiencies, we follow a four-step process. First, we refine the CUD dataset based on specific criteria by color experts. Second, we expand the input image information through pre-processing that is specialized for color deficiency vision. Third, we employ a multi-modality fusion architecture to combine features and process the expanded images. Finally, we propose a conjugate loss function based on the composition of the predicted image through the model to address one-to-many problems that arise from the dataset. Our approach is able to produce high-quality CUD images that maintain color and contrast stability. The code for CUD-Net is available on the GitHub repository

Figures (10)

• Figure 1: Comparisons of the non-CUD image, CUD-Net's predicted image, and CUD image. The top row is represented in normal vision, and the bottom row in deuteranopia vision. Numeral '5' appears on our predicted image in deuteranopia vision and complies with color preservation.
• Figure 2: Predicted image conversion for comparison experiments in the training set. The items are arranged in descending order of evaluation, starting from the left. It's noteworthy that all experiments were conducted using the identical dataset, which had been refined by color experts based on specific criteria.
• Figure 3: Ideal color conversion of the predicted image by balancing contrast and color preservation. The non-CUD object $a$ should widen the gap compared to the input image and preserve its original color harmony, whereas the CUD object after filter adjustment $\acute{b}$ preserves both contrast and color harmony.
• Figure 4: Overview of CUD image generation. The input is initially divided into three feature images prior to its passage through the CUD-Net. The CUD-Net, as depicted by the blue line in Figure \ref{['fig:04']}, proceeds to regress the node points of a piecewise linear function within the S and V channels. Ultimately, this filter is applied to the original input. On the right side of the predicted image, the content is distinguishable to individuals with both normal and deuteranopia vision. However, in the input image with deuteranopia vision, this discrimination is not feasible.
• Figure 5: Structural overview of CUD-Net.