Feature-attention module for context-aware image-to-image translation

In a summer2winter image-to-image translation, trees should be transformed from green to gray, but the colors of houses or girls should not be changed. However, current unsupervised one-to-one image translation techniques failed to focus the translation on individual objects. To tackle this issue, we propose a novel feature-attention module for capturing the mutual influences of various features, so as to automatically attend only to specific scene objects in unsupervised image-to-image translation. The proposed module can be integrated into different image translation networks and improve their context-aware translation ability. The qualitative and quantitative experiments on horse2zebra, apple2orange and summer2winter datasets based on DualGAN, CycleGAN and UNIT demonstrate a significant improvement in our proposed module over the state-of-the-art methods. In addition, the experiments on apple2orange dataset based on MUNIT and DRIT further indicate the effectiveness of FA module in multimodal translation tasks. We also show that the computation complexity of the proposed module is linear to the image size; moreover, the experiments on the day2night dataset prove that the proposed module is insensitive to the growth of image resolution. The source code and trained models are available at https://github.com/gaoyuainshuyi/fa.