Interleaved Zooming Network for Image Inpainting

Recently, deep learning-based techniques have shown great potential for image inpainting tasks thanks to the powerful representation ability of convolutional neural networks. However, current approaches failed to inpaint satisfactorily in both global structures and local textures, because they only use simple dilated convolution which hard to encode comprehensive deep features. In this paper, we propose Interleaved Zooming Network to tackle this issue. The network contains two branches in encode stage, major branch maintains the size of input features, zoom-out branch convolve and downsample the features by 4 times to perceive larger visual field of input signals. These two branches communicate in an interleaved manner, that is, each branch receive signals from previous layers in both branches. We empirically demonstrate that this interleaved zooming structure can help to generate better global structures and local details of missing regions of images. Compared with state-of-the-art deep learning-based inpainting methods, our approach provides more than 2dB gain on Places2 dataset with irregular mask.