CSAGAN: Channel and Spatial Attention-Guided Generative Adversarial Networks for Unsupervised Image-to-Image Translation

Unsupervised image-to-image translation is to learn a mapping function from one image domain to another with unpaired samples, which is an important task of computer vision. However, current unsupervised image-to-image translation methods only perform well on certain datasets. To handle the limitation, this paper proposes a novel framework termed as CSAGAN which contains a new discriminator structure, a novel attention module, and a new normalized function. The discriminator is an attention-guided feature pyramid discriminator which makes use of low-level and high-level features to determine an image's realness. The new attention module integrating channel attention and spatial attention can guide generators focus on the most discriminative regions of feature maps to generate high-quality translated images. Moreover, our attention module embedded into generators requires less computation compared with other self-attention methods. In addition, the new normalized function helps generators limberly control the variation of shape, color, and texture through learning parameters. Experimental results indicate that our approach performs better than the current state-of-the-art methods.