Conditional generative model for super-resolution of low-quality color fundus photographs

In the clinical diagnosis of eye diseases, color fundus photographs usually have a very large size to distinguish local fundus structure information or conduct combined evaluation with other modalities.However, current deep learning tasks usually require reducing the original images significantly to meet the requirements of graphics card memory during model training. Therefore, we use super-resolution technology to explore how to re-enlarge the small-sized images processed by the model back to the original size for diagnostic analysis. Specifically, it extracts high-frequency components during the enlargement of low-quality images to constrain the training process, which helps maintain the original physiological and pathological information and improve reconstructed image quality. Moreover, a carefully designed image degradation algorithm simulates real low-quality color fundus photographs. Comparison experiments show the model’s effectiveness in scale x2 and x4 tasks, and image denoising experiments prove its significant inhibitory effect on noise in the reconstruction process.