Image Inpainting Based On Conditional Generative Adversarial Networks

As one of the main investigation fields of image processing,image inpainting aims to fill the image defect caused by noise or improper storage,so that image restored could be consistent with the original one in texture and semantic and meet the requirement of visual perception.In recent years,enormous progress was achieved in this field with the development of deep learning.Benefiting from these achievements,restoring of scratch mark and small area defect was improved and complicated image inpainting such as large area or abnormity defect restoring become possible.Among several methods which use deep learning model to inpaint,weakness of reconstructing loss was existing of blur problem and weakness of adversarial loss was ignoring of influence of noise and losing the diversity of generated results of GANs.By experiments and analyzing,we contribute these problems to the immanent ill-posed of image inpainting.Hence,we consider that adding extra information,which will reduce one-to-many mapping and improve maneuverability will be an ideal way to improve algorithm.To develop form of guiding information,we proposed a new method which could guide the restoring process by comparing the image to be restored with similar ones.We first tried channel-wise connection network architecture to repair images,then analyzed the differences between guiding information and the origin image and proposed a double-encoder network.This network will extract information from both guiding image and image to be restored and use different network architectures and initialization methods in a more targeted way.Finally,double-encoder network and guidance image proved their validity in improving image inpainting by experiment.