Fusing dense and ReZero residual networks for super-resolution of retinal images

Abstract

• A deep neural network (DRRN) for super-resolution of retinal image is built. • Use dense connections in the feature mapping to reuse feature information. • Exploit local and global residuals learning to recover retinal image details. • Use ReZero to dynamically learn the weight of the residual. Retinal images have important reference value for the analysis of cardiovascular diseases such as diabetes and hypertension. However, because of the large number of artifacts at the end of the blood vessel and the unclearness of the intersection of the blood vessel, it will cause the inaccuracy of the retinal blood vessel segmentation, thereby affecting the diagnosis of retinal diseases. To solve this problem, we proposed dense and ReZero residual networks for super-resolution of retinal images (DRRN). Firstly, we used a dense connection module to obtain more feature information. Then, we added local residual learning and global residual learning respectively to improve the resolution of retinal image blood vessels from coarse to fine. We also introduced ReZero (residual with zero initialization) to the local residual learning, which can dynamically facilitate well-behaved gradients in a deep network with dense connection modules to make the network converge faster. Finally, we used a tandem structure to reconstruct the local details of the retinal image. The results showed that our method can better reconstruct the details of the retinal image, and can accelerate the processing speed on STARE, DRIVE, and CHASE_DB1 datasets.