An Unsupervised Fundus Image Enhancement Method with Multi-Scale Transformer and Unreferenced Loss

Color fundus images are now widely used in computer-aided analysis systems for ophthalmic diseases. However, imaging can be affected by human, environmental, and equipment factors, which may result low-quality images. Such quality will interfere with diagnosis. Existing methods enhancing focus more on the overall visualization of image rather than capturing pathological structural features at finer scales sufficiently. In this paper, we design an unsupervised method that integrates a multi-scale feature fusion transformer unreferenced loss function. Due to microscale caused unpaired training, construct Global Feature Extraction Module (GFEM), combination convolution blocks residual Swin Transformer modules, achieve extraction information different levels while reducing computational costs. To improve blurring details deep networks, define functions model’s ability suppress edge sharpness degradation. addition, uneven light distribution also affect quality, so use priori luminance-based attention mechanism illumination unevenness. On public dataset, improvement 0.88 dB PSNR 0.024 SSIM compared state-of-the-art methods. Experiment results show our outperforms other learning terms vascular continuity preservation fine features. framework have potential medical applications.