Runtong Zhang’s research while affiliated with University of Electronic Science and Technology of China and other places

Unsupervised image-to-image translation such as CycleGAN has received considerable attention in recent research. However, when handling large images, the quality of generated images are not in good quality. Progressive Growing GAN has proved that progressively growing of GANs could generate high pixels images. However, if we simply combine PG-method and CycleGAN, it must bring model collapse. In this paper, motivated from skip connection, we propose Progressive Growing CycleGAN (PG-Att-CycleGAN), which can stably grow the input size of both the generator and discriminator progressively from $256\times 256$ to $512\times 512$ and finally $1024\times 1024$ using the weight ${\alpha }$. The whole process makes generated images clearer and stabilizes training of the network. In addition, our new generator and discriminator cannot only make the domain transfer more natural, but also increase the stability of training by using the attention block. Finally, through our model, we can process high scale images with good qualities. We use VGG16 network to evaluate domain transfer ability.

Images generated from Cycle-Consistent Adversarial Network (CycleGAN) become blurry especially in areas with complex edges because of loss of edge information in downsampling of encoders. To solve this problem, we design a new model called ED-CycleGAN based on original CycleGAN. The key idea is using a pre-trained encoder: training an Encoder-Decoder Block (ED-Block) at first in order to get a difference map, which we call an edge map and is produced by the subtraction of input and output of the block. Then, the encoder part of a generator in CycleGAN share the parameters with the trained encoder of ED-Block and they will be frozen during training. Finally, by adding the output from a generator to the edge map, higher quality images can be produced. This structure performs excellently on “Apple2Orange”, “Summer2Winter” and “blond-hair2brown-hair” datasets. We use SSIM and PSNR to evaluate resolution of results and our method achieved the highest evaluation scores among CycleGAN, Unit and DiscoGAN.