An Improved Res-UperNet Method for Retinal Feature Segmentation

Aiming at the problems of low detection accuracy and high influence by image quality of existing semantic segmentation algorithms in retinal feature segmentation, this paper proposes an improved Res-UperNet method for retinal feature segmentation (RURFS). In this paper, under the framework of the UperNet network model, an Resnet backbone network structure is introduced to effectively stack the number of neural network layers and thus improve the information extraction ability of the network; the number of network layers of the backbone network is increased to enable it to extract higher-level information; meanwhile, a data enhancement strategy of random rotation is used to enable the network to learn features in different directions. The experimental results show that the average intersection degree (mIou) of RURFS is improved by about 0.7
^%
compared with the original model on CHASE _ DBl dataset, and for different retinal features, the network can segment the features better at different luminance, and the improved network has better robustness to complete the segmentation of retinal features with strong segmentation capability.