Multiscale Feature Extraction and Attention Mechanism Generative Adversarial Network for Super-Resolution and Deblurring of Fundus Images

High-resolution and clear fundus images are essential to help physicians diagnose lesions. However, the imaging quality of acquired fundus images often has errors due to differences in operator experience and equipment limitations. To address this problem, this paper proposes a super-resolution network for retinal fundus images based on generative adversarial networks (GANs). The network aims to improve the resolution of fundus images and restore fine retinal structure and lesion details. First, based on the analysis of the ophthalmic mirror system, we designed a new degradation model to simulate the effects of various unfavorable factors on fundus images, and thus constructed a batch of fundus image datasets for fundus image super-resolution work. In order to enhance the network's ability to extract local information, we introduced a texture reply block based on coordinate attention. Meanwhile, in order to capture the fundus image features at different scales, we also add a multi-scale feature extraction block to realize the fusion of multi-scale features. Experimental results show that our network is able to reconstruct high-quality fundus images, and the proposed method outperforms other super-resolution deblurring methods in both PSNR and SSIM metrics. This result provides strong support for accurate diagnosis of fundus images.