Innovative paths of CycleGAN loss functions and evolution of cross-domain applications

School of Advanced Technology, Xi’an Jiaotong-Liverpool University, Suzhou, 215028, China

^* Corresponding author: Lijia.Liu22@student.xjtlu.edu.cn

Abstract

Image translation is one of the most popular topics in computer vision. The development of CycleGAN addresses several issues in multi- task applications, including image translation. This study systematically reviews the innovative design of CycleGAN loss functions and their cross- domain applications in medical and industrial imaging. First, to address the limitations of basic CycleGAN, the core innovations in loss functions are provided: the perceptual loss design overcomes the constraints of pixel-level matching (such as MSE and L1 loss) by aligning high-level semantic features. Identity loss restrains mode collapse through imposing identity mapping. Domain-specific loss customizes constraints in specific scenarios, and the combined loss function adjusts the weights of multi-objective tasks to achieve balance. Second, for cross-domain applications, based on specific tasks in medical image synthesis and enhancement, industry defect detection, and image dehazing. Comparing the performance between the improved CycleGAN, basic CycleGAN, and other methods. It shows that improved CycleGAN with innovative loss function has significant advance. As a result, the innovation of CycleGAN loss functions is not only a technical breakthrough in generative adversarial networks but also provides a problem-driven design paradigm for cross-domain image processing. It is highly effective in cross-domain application design.