We propose a new scheme for EPI distortion correction, which implements unsupervised learning, trained with readily available images, such as ImageNet2012 dataset. The distortion-corrected image is obtained by the MR image generation function using the input distorted images and the frequency-shift maps that are the outputs of the network. Two distorted images obtained with dual-polarity phase-encoding gradients are the inputs of the neural network. The neural network estimates the frequency-shift maps from the distorted images. To train the neural network, unsupervised learning was conducted by minimizing the L1 loss between input distorted images and the estimated distorted images.