<p>High-quality retinal fundus images are crucial for accurate ophthalmic screening and diagnosis, but imaging interferences often lead to degradation of fundus image quality. Thus, fundus image enhancement is worth studying. Retaining retinal structural information to improve the enhancement performance is a major challenge for fundus image enhancement models, especially in the case of no supervision from paired images. Furthermore, the inability to remove granular artifacts from fundus images remains a problem for existing unsupervised enhancement models. To solve these problems, we propose a CycleGAN-based unsupervised fundus image enhancement network with high-frequency feature fusion and artifact processing in this paper. Specifically, we extract the high-frequency component of an input fundus image and encode it directly into the generator to retain retinal structural details. Moreover, we develop an artifact processing module based on pix2pix to improve the visual quality of enhanced fundus images. We also integrate an attention mechanism into the feature fusion process to effectively capture global information and strengthen features. Last but not least, we present the high-frequency consistency loss to constraint the enhancement network for keeping the fundus biological structures more effectively. Extensive experiments conducted on EyeQ, Drive, and RF data sets demonstrate that our model outperforms state-of-the-art ones both quantitatively and qualitatively, and can benefit the downstream tasks of diabetic retinopathy grading and retinal vessel segmentation.</p>