Generative image completion with image-to-image translation

Though many methods have been proposed, image completion still remains challenge; besides textured patterns completion, it often requires high-level understanding of scenes and objects being completed. More recently, deep convolutional generative adversarial networks have been turned into an efficient tool for image completion. Manually specified transformation methods are having been replaced with training neural nets. Hand-engineered loss calculations for training the generator are replaced by the loss function provided by the discriminator. With existing deep learning-based approaches, image completion results in high quality but may still lack high-level feature details or contain artificial appearance. In our completion architecture, we leverage a fully convolutional generator with two subnetworks as our basic completion approach and divide the problem into two steps: The first subnetwork generates the outline of a completed image in a new domain, and the second subnetwork translates the outline to a visually realistic output with image-to-image translation. The feedforward fully convolutional network can complete images with holes of any size at any location. We compare our method with several existing ones on representative datasets such as CelebA, ImageNet, Places2 and CMP Facade. The evaluations demonstrate that our model significantly improves the completion results.