Segmentation of text regions in ancient inscriptions via advanced text segmentation methodologies is crucial for the restoration of textual structure and the decipherment of historical content. Nonetheless, the dearth of annotated datasets specific to ancient text rubbing segmentation, compounded by the prohibitive expense associated with such annotations, poses challenges to the application of conventional supervised learning models in this domain. In addressing these challenges, we introduce CycleTsGAN, an unsupervised learning algorithm tailored for ancient text rubbing segmentation. By recasting the segmentation task as a style transfer problem, CycleTsGAN ingeniously circumvents the necessity for annotated data, thereby overcoming a key bottleneck in the field. In addition, in order to meet the requirements of model training, JinwenSeg, JinwenSeg-1k and JinwenSeg-Eval datasets were constructed. Experiments have shown that the CycleTsGAN model can generate clear and accurate segmentation results in the task of ancient text rubbings segmentation.