محسن رمضانی

Restoring damaged region in digital images (i.e., image inpainting) can be considered as a difficult problem which gets proportionally harder based on the severity of the damage. In the last few years, there is a progress in tackling this issue through using deep learning models. In this study, according to the successful applications of GANs in different fields, a new approach is presented for image inpainting. The proposed algorithm contains a generator and a global discriminator. The generator is responsible for recovering the missing area, and the global discriminator relates to identifying whether the repair area is correct or not. The architecture of the generator consists of two auto-encoder. Moreover, Wasserstein GAN loss is used to ensure the stability of training. As input image a 32 by 32 icon image is also used to semantically guide the generator, and then concatenating with the corrupted image for filling the lost part or regions without losing some existing objects or predicting unwanted objects or shapes. The guide image can be proposed by the user of application or some other cases such as watermarking can be considered. This method is qualitatively and quantitatively compared to the state-of-the-art models which use a Generative Adversarial Network. These approaches can produce convincing visual structures and textures, but they frequently produce deformed structures, blurry textures or loss objects that are out of sync with the surrounding areas. The presented results on CelebA-HQ dataset demonstrate that the proposed model can deal with large-scale missing pixels and generate realistic results.