Cycle-GAN based Face Aging Method

In this study, we try to explore the possibility of cycle-GAN in the field of face aging and the effect of cycle-GAN based aging model. The reason why we used cycle-GAN for this study is that cycle-GAN does not require paired data as a training set. Paired data is expensive and difficult to obtain paired data. Therefore, if unpaired data can be used as a training set, it will greatly reduce the cost of facial aging. We plan to use PSNR and SSIM to measure the truthfulness of cycle-GAN generated images. We then quantified the aging effect using a CNN-based DEX network to assess the age of the aging images. In the process of research, we found that the model was unbalanced in processing images of different genders and ethnic groups, so we tried to make the model more balanced in the above cases through classification training and pre-training, which achieved good results. Finally, we developed a GUI as a convenient interface for users to use cycle- GAN model to do face aging. Through a series of quantitative analyses, we were able to conclude that our model has a good effect in the field of facial aging and is superior to other aging models based on deep learning. We also provide the code of this project on GitHub: https://github.com/MeditatorE/Face-time-travel-machine