Unsupervised Transfer Learning for Generative Image Inpainting with Adversarial Edge Learning

Abstract

Deep learning-based image restoration techniques have made great progress in recent years, and EdgeConnect network has achieved good results in image restoration. We find that EdgeConnect suffers from a complex training process and poor migratability, which reduces its usability in practical applications. We explore the reasons for the poor transferability learning and generalization of the EdgeConnect model, and propose a small-sample unsupervised joint transfer learning method for the case of small datasets and low data similarity. The method combines a large amount of Fine-tune with a small amount of direct migration training to enable the network to learn new knowledge of the target domain while avoiding overfitting and negative migration. We perform migration learning and evaluation on 600 images from Paris StreetView with a pre-trained model obtained on the CelebA dataset, and show that it outperforms other current methods in terms of quality.