Enhancing Ishihara and educational images using machine learning: toward accessible learning for colorblind individuals

TL;DR

This paper uses machine learning to improve Ishihara and educational images for people with red-green color blindness, making them more accessible.

Contribution

A novel daltonization method with optimized enhancement strength is proposed to improve image accessibility for colorblind individuals.

Findings

Optimal enhancement strength (α) of 0.54 for deuteranopia and 0.64 for protanopia achieved high contrast gains with minimal distortion.

OvA strategy reached 99.7% accuracy in classifying simulated and reference images.

Survey results showed enhanced images improved recognition of digits and symbols for colorblind students.

Abstract

Color Vision Deficiency (CVD) affects over 300 million individuals worldwide, with protanopia and deuteranopia being the most common subtypes, causing red–green confusion. This study leverages machine learning to (a) classify reference (considered as normal vision) and simulated protanopia and deuteranopia Ishihara plate images, (b) generate corresponding enhanced versions of these images, and (c) provide improved textbook diagrams (from NCERT books) and other pseudochromatic figures for CVD students, validated through feedback from diagnosed individuals. Tritanopia and milder forms of CVD were excluded in this study. A dataset of 1,400 Ishihara plates was processed to simulate protanopia and deuteranopia perception via standard Red Green Blue (sRGB) to long-, medium-, and short-wavelength cone (LMS) modeling. Enhanced images were generated using a daltonization function defined by the error between reference and simulated images, with enhancement strength (α) optimized to maximize contrast gain while minimizing distortion. Feature embeddings from ResNet-50, EfficientNet-B0, and DenseNet-201 were fused and reduced via PCA, followed by One-vs-All (OvA) (classifiers: linear support vector machine, logistic regression, and decision tree), random forest, gradient boosting, and neural network. Results showed optimal enhancement at α = 0.54 for deuteranopia and 0.64 for protanopia, achieving contrast gains of 69.6 and 64.3, respectively, with minimal color distortion (ΔE ≈ 4.9) and negligible clipping (<0.002). The OvA strategy achieved 99.7% accuracy, while MLP reached 100% across metrics. Surveys with 15 diagnosed students confirmed substantial perceptual improvement: recognition of previously unreadable digits and symbols increased from <20% to full visibility, with mean ratings above 4/5 for enhanced images. The OvA technique integrated with daltonization can assist in enhancing Ishihara and educational images in real time.

Figures