<p>Rubbing images of oracle bones are crucial for studying ancient Chinese culture, but their quality often hinders character recognition and analysis. Current super-resolution approaches face notable limitations when applied to oracle bone rubbing images, particularly in preserving the integrity of genuine character structures, and often introduce character distortions and visual artifacts. To address this issue, we propose an unsupervised super-resolution method for oracle bone rubbing images, named OBISR. The model is based on a generative adversarial network and introduces a two-stage reconstruction generator built upon Transformer architecture, along with a more stable generator variant obtained via the Exponential Moving Average (EMA) mechanism. Meanwhile, an artifact loss function is employed to suppress artifacts and distortions in character regions, and a set of local discriminators is designed to enhance detail perception. Experiments show OBISR outperforms existing methods across quantitative metrics and visual evaluations, offering superior enhancement for oracle bone rubbing images.</p>