基于方向条件的循环一致性生成对抗网络

Cycle-consistent generative adversarial network based on directional conditions

为了充分利用非配对数据进行图像翻译、减少网络参数和提高训练速度,采用条件生成对抗的监督训练与循环一致性生成对抗的无监督训练相结合的方法,设计了基于方向条件对偶的生成网络,同时采用Patch结构的判别器输出多维判别结果,结合感知损失和同一损失与循环一致损失,设计了更有效的损失函数。通过在相同数据集上与CycleGAN进行对比实验,验证了所提网络在非配对图像翻译任务上,网络参数减少34%,生成图像的PSNR值平均提升4.9%,SSIM值平均提升6.3%,并且有效提升了训练速度和重建图像的质量。

In order to make full use of unpaired data for image translation training,reducingnetwork parameters,and improving training speed a combination of supervised training with conditional generative adversarial and unsupervised training with cyclic-consistent generative adversarial is used to design a generative network based on directional conditional pairwise,while a discriminator with a Patch structure is used to output multidimensional discriminant results,and a more efficient loss function is designed by combining perceptual loss and same loss with cycle-consistent loss.By comparing the experiments with CycleGAN on the same dataset,the network parameters of the proposed network in this paper are reduced by 34%,the PSNR values of the generated images are improved by 4.9%on average,the SSIM values are improved by 6.3%on average on the non-pairwise image translation task and the training speed and reconstructed image quality are effectively improved.