Image Inpainting with Pre-Trained Deep Neural Networks

Abstract

Images are exposed to deterioration over years due to many factors. These factors may include environmental factors, chemical processing, improper storage, etc. Image inpainting has gained significant attention from researchers to recover the missing regions in images. In this thesis, four new image inpainting algorithms using Deep Convolution Neural Networks (CNN) are proposed. In the first algorithm, a new methodology for minimizing the blurriness of images when only using L2 loss is proposed. In the second algorithm, we succeeded in the inpainting of missing regions located in various positions across images using GAN, rather than inpainting of a fixed center missing region across images. In the third algorithm, a new training methodology employing dilated convolution with a mask modification step and a global discriminator are proposed for the inpainting of a random-shaped missing region across images. In the fourth algorithm, we proposed a novel based algorithm for face inpainting, we changed the training methodology to begin with a guidance network, trained to hallucinate the missing structures then using these hallucinations to guide the inpainting process afterwards. We trained and validated our models using benchmark imaging datasets. All performed experiments have shown that the proposed algorithms have better performance in recovering missing regions in images over other state-of-art algorithms. This was done at the expense of the new structure and long training time. Keywords: Image inpainting, GAN, L1 reconstruction loss, L2 reconstruction loss, GAN loss, adversarial loss, convolution neural network, Encoder-Decoder, and dilated convolution.