Capsule based image translation network

Along with the development of deep learning, the research of computer vision has attracted much attention in recent years. Among various research topics, image translation is quite a hot one, which aims to translate an image into another domain according to the exact application purpose. In this paper, we present a novel capsule based conditional generative adversarial network that can automatically synthesize the image with realistic and aesthetically pleasing effect from a given image. By designing the capsule blocks in both the generator and discriminator and a novel multi-way optimization function, our framework can generate more realistic images than other frameworks. For the specific application, effect rendering, an effective line preservation loss is well designed and helps the framework to focus the synthesis of the texture around the lines, which not only improves the local line visual effect, but also enhance the global lighting effect around the lines. The experiments show that our improvement is effective and provides significant help to the overall framework. Our framework is tested on the effect rendering dataset and we compare our approach with multiple state-of-the-art methods. We achieved the best results.