Unsupervised image-to-image translation by semantics consistency and self-attention

Unsupervised image-to-image translation is a challenging task for computer vision. The goal of image translation is to learn a mapping between two domains, without corresponding image pairs. Many previous works only focused on image-level translation but ignored image features processing, which led to a certain semantics loss, such as the changes of the background of the generated image, partial transformation, and so on. In this work, we propose a method of image-to-image translation based on generative adversarial nets(GANs). We use autoencoder structure to extract image features in the generator and add semantic consistency loss on extracted features to maintain the semantic consistency of the generated image. Self-attention mechanism at the end of generator is used to obtain long-distance dependency in image. At the same time, as expanding the convolution receptive field, the quality of the generated image is enhanced. Quantitative experiment shows that our method significantly outperforms previous works. Especially on images with obvious foreground, our model shows an impressive improvement.