UMGAN: Generative adversarial network for image unmosaicing using perceptual loss

Image mosaicing conceals sensitive parts of an image. The objective of this work is to recover hidden semantic structure under mosaiced parts, especially focusing on facial images. While recent image completion methods based on deep learning have shown promising results on recovering damaged parts in an image, they have not addressed the problem of image unmosaicing. We present a Generative Adversarial Network (GAN) approach to image unmosaicing called UMGAN, which is an image-to-image translation method. We have found that exploiting perceptual loss together with low level $l-{1}$ loss and high level Structural SIMilarity (SSIM) loss is quite effective to attain visually plausible and semantically consistent results. We have evaluated our method on the CelebA and MIT-CBCL image datasets and achieved better perceptual results than state-of-the-art image completion methods.