Image Style Transfer Based on DPN-CycleGAN
Mingyu Yang, Jianjun He · 2021
Image style transfer has always been a hot topic in the field of image generation. Generative Adversarial Network (GAN) is used to realize image style transfer can greatly reduce the workload, and get more fruitful results. The traditional image style transfer algorithm needs to convert between two paired images, but the paired data set used for training is difficult to obtain. In order to avoid being limited by data sets and improve the efficiency of image style transfer, this paper proposes an improved Cycle-Consistent Adversarial Network (DPN-CycleGAN), which uses Dual Path Network (DPN) replaces the Deep Residual Network (ResNet) of the original network generator, and adds Identity Loss on the basis of the original loss. The changes make the network structure more simplified, reduce the computational complexity, improve the network performance to a certain extent, and improve the quality of the image generated by style transfer. Finally, the experimental results show that the SSIM value and PSNR value of the generated images are increased by 3.6% and 9.7% on average, which proves the effectiveness of the DPN-CycleGAN image style transfer algorithm proposed by this paper.