Effect Analysis of Image Translation by Controlling Size of Dataset and Adjusting Parameters for CycleGAN
Keang Xu, Ziyu Zhou, Lesi Wei · 2021
Image translation can be achieved effectively by CycleGAN. However, it is not clear whether the dataset size and parameter adjustment would influence the performance of CycleGAN. Therefore, this paper uses the different kinds of dataset and different dataset sizes to evaluate the performance of the model. Furthermore, parameters are adjusted in CycleGAN to further evaluate the performance of the CycleGAN-based method for image translation. The experimental results show that with the increases of the size of the dataset, except for some special situations, the loss of the model will decrease when different loss shows different descending amplitude which means small size of dataset would lead to overfitting. In addition, this cycle GAN model is more suitable for the “horse to zebra” dataset, because this dataset has the least loss. Cycle consistency or its weight, hyper parameters adjusting can achieve the better unpaired image resolution than the original image, which shows this model can do a fairly good translation. Our research is of great significance to explore the best parameter settings and dataset size of image translation optimization based on CycleGAN.