Image Style Transfer Based on CycleGAN
Lisha Yao, Qiaoqiao Feng · 2023
This article uses Cycle Generative Adversarial Networks, which uses Cycle GAN to transfer the style of images and convert natural images into images with a certain style. The CycleGAN algorithm can effectively solve the problem of not being able to complete image style conversion when images are not paired, and by adding cyclic consistency loss, CycleGAN solves the problem of insufficient data distribution capture ability. This article verifies on horse2zebra data that compared to the basic model, skip connections solve the problem of background color distortion and converge quickly. In the transition task from horse to zebra and from zebra to horse, the pass rate of the artificial perception test is 29%. This testing metric represents the quantitative results of subjective testing, but there is currently no widely used evaluation metric for instance object style transfer.