Papers Address:
http://iizuka.cs.tsukuba.ac.jp/projects/completion/en/
aims:
Filling an image, the deleted region of an arbitrary shape to complete the filling images of any resolution.
Network configuration:
Complete the network: the network is fully completed convolution, used to repair the image.
Discriminator global context: to complete the image as input, the scene recognition global consistency.
Local context discriminator: only a small region of interest around the complete area, in more detail to determine the quality of appearance.
Image complete the network is trained to deceive two contexts discriminator network, which requires it to generate overall consistency and detail in the image indistinguishable from the real image.
Experimental results:
