Image Correction for Improving Visual Acuity Using Zernike-Based Vision Simulation

Abstract

This paper proposes novel image correction method to improve visions of individuals with refractive errors such as myopia, hyperopia, and astigmatism. Refractive errors can be corrected by eyeglasses, contact lenses, or surgery, but there are problems such as the need for maintenance or cost. Our method corrects images so that the corrected images are perceived similarly to the original images without eyeglasses or contact lenses. Our method consists of two models: image correction model, and blur simulation model. The image correction model is a convolutional neural network model designed for super resolution tasks, and the blur simulation model simulates perceived blur of a human eye. Perceived blur can be modeled by Zernike polynomials-based point spread function (PSF), and blurred images are computed by convolving the images with PSF. Our model works as follows. First, images are fed to image correction model to produce corrected images. Then, these corrected images are passed to blur simulation model. Image correction model is trained so that blurred corrected images become identical to the original images. After training is completed, only first step is required to correct images. Our approach can correct any types of images using the same model without major contrast loss.