Efficient restoration of low-quality fundus images based on prior knowledge using a diffusion model

Fundus images are important tools for diagnosing and detecting ocular diseases in clinical practice. However, the quality of fundus images is often uneven due to differences in sensor performance, lighting conditions, and patient cooperation. Low- quality fundus images can pose diagnostic challenges for physicians. Diffusion models have achieved remarkable results in the field of image restoration, but their training process consumes a lot of resources, so it is difficult to efficiently train a feasible model. In order to improve the quality of low-quality fundus images, this study proposed an efficient restoration method based on a prior knowledge diffusion model. This method utilizes a pre-trained prior knowledge extraction network to guide the training process of the diffusion model, thereby improving the training efficiency and enhancing the ability to generate high-quality images. In addition, a dual-branch prior knowledge texture alignment network is designed to better extract prior knowledge. Through knowledge distillation, the low-quality image prior knowledge extraction network (L-CPEN) is aligned with the high-quality image prior knowledge extraction network (G-CPEN). A residual block is introduced in the L-CPEN network to better learn and align the differences between the original image and the low-quality image. At the same time, the texture alignment method helps preserve clinically relevant important information in fundus images. Experiments based on the EyeQ dataset show that this method significantly improves the restoration effect of low-quality fundus images and improves the PSNR and SSIM values.