Camera-in-the-loop システムによる視覚再現モデルを用いた屈折異常者向けの画像補正法

抄録

In this paper, we propose an image correction method to improve visions of individuals with refractive errors such as myopia, hyperopia, and astigmatism. Refractive errors can be corrected by eyeglasses or contact lenses. However, there are problems such as maintenance or cost. Our method corrects images so that the corrected images are perceived similarly to the original images without wearing eyeglasses. Two convolutional neural network models are employed in our system: blur simulation model and image correction model. Blur simulation model is trained to simulate optical blur using a blur-induced camera, and image correction model produces images that can be perceived similar to the original images through the blur simulation. In our method, blur simulation model learns how corrected images are perceived while training image correction model, in addition to pretraining of clean-to-blurry images conversion to avoid implausible blur simulation for corrected images. Our method is fully software-based on inference, thus we do not require any special equipment except for computers. The images produced by our model provide better vision without significant contrast loss.