Loss functions for Style Transfer with CycleGAN

Generative Adversarial Networks has been used in many fields now, and it is particularly essential in the field of computer vision. In respect of image-to-image translation, CycleGAN is an important part. In this paper, CycleGAN is used to translate portrait photographs to sketches, and ℓ1 loss, ℓ2 loss, perceptual loss and their combination losses are compared to find a cycle consistency loss function with better performance. After evaluating the generated image quality through two metrics: the peak signal to noise ratio (PSNR) and structural similarity (SSIM), the conclusion that combining the perceptual loss with ℓ1 and ℓ2 makes the generated images more structurally similar with their reference image is drawn.