We present a novel adversarial framework to solve the arbitrarily sized image random inpainting problem, where a pair of convolution generator and discriminator is trained jointly to fill the relatively large but random "holes". The generator is a symmetric encoder-decoder just like an hourglass but with added skip connections. The skip connections act like information shortcut to transfer some necessary details that discarded by the "bottleneck" layer. Our discriminator is trained to distinguish whether an image is natural or not and find out the hidden holes from a reconstructed image. A combination of a standard pixel-wise L2 loss and an adversarial loss is used to guided the generator to preserve the known part of the origin image and fills the missing part with plausible result. Our experiment is conducted on over 1.24M images with uniformly random 25% missing part. We found the generator is good at capturing structure context and performs well in arbitrary size images without complex texture.
展开▼
