Image Restoration Using Deep Learning Based Image Completion

Abstract

Digital images can experience various disturbances in acquisition and storage, one of which is a disturbance indicated by damage to certain areas of the image field and causes the loss of some of the information represented by the image. One of the ways to restore an image experiencing disturbances like this is with image completion technology. Image completion is an image restoration technology capable of filling in or completing missing or corrupted parts of an image. Various methods have been developed for this image completion, starting from those based on basic image processing to the latest relying on artificial intelligence algorithms. This study aims to develop and implement an image completion model based on deep learning with the transfer learning method from the completion.net architecture. Using the Facesrub training dataset consisting of a collection of unique facial photos allows the model to understand facial attributes better. Compared to conventional image completion based on image patches, the method developed in this study can perform image filling in image gaps with more realistic results. Based on visual tests conducted on respondents, the results obtained enable respondents to understand all the information represented by the restored image, similar to the original image.