WGAN Based Edge Preserved De-Blurring Using Perceptual Style Similarity
Minsoo Hong, Yoonsik Choe · 2018
Generative Adversarial Network (GAN) is an effective generative model and can be used for de-blurring. In this paper, we propose an edge preserved de-blurring method using Wasserstein generative adversarial network with gradient penalty (WGAN-GP), which is based on conditional GAN. Also, since detailed-edge is the most important factor in de-blurred image, in order to preserve detailed-edge and capture its perceptual similarity, the style loss function is added to represent the perceptual information of the edge. Consequently, the proposed method improves the similarity between sharp images and blurred images by minimizing Wasserstein distance, and well captures the perceptual similarity using style loss function, considering the correlation of features in Convolutional Neural Network. Experiments depict that the proposed method achieves 0.98 in SSIM that is higher performance, compared to other conventional methods such as filter based methods and content based method.