Unsupervised image-to-image translation can map local textures between two domains, but typically fails when the domain requires big shape changes. It is difficult to learn how to make such big change using the basic convolution layer, and furthermore it takes much time to learn. For faster learning and high-quality image generation, we propose to use Cycle GAN that is combined with Resnet in a network that is connected with the residual block for upsampling to make big shape change and construct faster image-to-image translation.
In unsupervised image-to-image translation, when the shape of data images is very different, the encoderdecoder format is used for learning. However, there is a problem that it takes some time to learn because there is no influence between layers. The main idea of Resnet's idea is to use a residual block that creates a skip connection so that the gradient can flow well. This is similar to the Long Term Short Memory (LSTM), which introduces a forget gate and so on to better flow the gradient of the previous step. In this study, by using these residual blocks in the generator upsampling, it is possible to flow the big feature well, but it does not have big shape transformation, and it is easy to learn because the image quality is improved, the network is structured with high speed. and a generator with residual blocks is constructed. The architecture of this study is based on DiscoGAN and CycleGAN. The existing DiscoGAN has the basic encoderdecoder structure as shown in Figure 5, and it is easy to change the format. However, since it has the basic decoder structure of the network, it is difficult to learn upsampling through the narrow bottleneck layer and output image size. And it has a low resolution limit due to the low capacity of the network. Therefore, the generator in this study includes residual bocks in the upsampling layer, thus making it easier to learn and be capable of producing images of high quality.
| Our | DiscoGAN | |
|---|---|---|
| 500 Epoch Time | 41:54:60 | 61:41:44 |
| Classification Error | 8.92% | 8.77% |
In this study, we used a residual block in the upsampling structure in the generator to maintain the shape transformation and construct a network that takes much less time to generate high quality images. One of the problems in existing unsupervised image-to-image translation, which takes a long learning time, is solved by using a residual block, and the image can be converted at a higher image quality than the existing research. In future work, we plan to construct a more efficient network by combining new ideas from generator or loss function as well as generator structure.