Extraction and Parallelism of U-Net Network Improvement Algorithm in Retinal Vessels

Xiaobo Qu
College of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, Shanghai, China
Su Yu
College of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, Shanghai, China

Abstract:

A U-Net retinal vascular image acquisition solid model with residual blocks, circular convolution module and spatial channel extrusion encouragement module is proposed to address the poor segmentation results caused by fine retinal vessels, small detail feature loss, gradient descent method and occurrence of explosion, and the parallelism is scientifically studied. First, a series of arbitrary enhancements are applied to expand the test set and prepare the processed data, and then the residual blocks are introduced into the U-Net solid model to prevent the segmentation precision from saturating and then declining rapidly with the increase of network depth and to enhance the accounting cost. The bottom layer of the U-Net network is replaced by a circular convolution module to obtain the underlying features of the image, accumulate precipitated features to enhance the lexical information between the preceding and following texts, and obtain a more efficient segmentation entity model; finally, the indoor space wireless channel reduction encouragement module is placed in the middle of the convolution layer. The network model focuses on this channel according to the channel with better exploration features, narrows down the irrelevant channels, strengthens the learning of important lexical-semantic feature information courses, learns efficient feature information using the whole process of practice, and improves interference resistance. In addition, this study uses the unique network architecture of U-Net Internet to accelerate parallelism by increasing the parallelism of logical control components. Ultimately, the segmentation and extraction method in the paper has better segmentation practical results when compared with other network segmentation and extraction methods.