Ishihara-Like MNIST Dataset

Color blindness, a condition that affects approximately 8% of men and 0.5% of women worldwide, hinders individuals from perceiving certain colors accurately. One of the most common tests for diagnosing color blindness is the Ishihara color blindness test, which uses a series of plates containing numerals obscured by dots of various colors. The test relies on the color contrast between the numeral and the background, designed to exploit the vision deficiency and reveal a person’s inability to distinguish the hidden numbers.

Inspired by this concept, we developed the Ishihara-Like MNIST dataset. This dataset is generated by applying the principles of the Ishihara test to the widely known MNIST dataset of handwritten digits. Each image in the dataset contains a numeral from MNIST that has been creatively obscured with colored dots, closely mimicking the structure of an Ishihara test plate. The result is a set of images where the digits are embedded in a visually complex background, challenging a machine learning model’s ability to recognize and classify them.