Object Removing via Multi-Column Gate Convolutional Neutral Network

Abstract

Object removal is to remove unwanted objects in a photograph. It can be realized by the inpainting technique with irregular mask. Generative Image Inpainting with Contextual Attention [n] and Partial convolution cannot work well with irregular mask，so we present a Multi-column Gate Convolutional Neutral Network. It can work well with both regular and irregular mask. Our model adopts gate convolution which is more suitable in inpainting task than traditional convolution. Special context loss, semantics loss and spatial reconstruction loss are used for better performance. We train the model with CelebA and Landscape. A comparison is made between our model and some models published on CVPR etc, our model’s results can generate more realistic images than existing ones for object removal.