摘要
为提升眼底视网膜血管分割的准确度,文中算法在U-Net基础上进行改进,提出了一种基于全卷积网络的眼底视网膜血管分割算法。将原始U-Net网络上下采样中的卷积层改为Inception模块,该模块可以在平衡网络深度和宽度的前提下,增大上下采样过程中特征向量的多尺度信息量与形状结构形态量;将原始网络底部的池化层部分改为金字塔型空洞卷积,该模块可以针对细小血管特征较难捕捉、分叉处丢失信息量较大的问题,增大局部感受野,更好地完成眼底视网膜血管分割技术。该算法在DRIVE数据库的准确率为95.39%,敏感度为78.42%,F1值为0.8213,较现有的先进算法有一定的提升。
In order to improve the accuracy of fundus retinal vascular segmentation,the algorithm in this paper is improved on the basis of U-Net,and a new algorithm of fundus retinal vascular segmentation based on full convolutional network is proposed.The convolutional layer in the original U-Net network up-and-down sampling is changed into Inception module,which can increase the multi-scale information and shape,structure and morphological quantity of the feature vectors in the up-and-down sampling process on the premise of balancing the depth and width of the network.The original pooling layer at the bottom of the network was changed into pyramidal cavity convolution.This module can solve the problems of small blood vessels that are difficult to capture and the information lost at the bifurcation is large,increase the local receptive field,and better complete the fundus retinal vessel segmentation technology.The accuracy of this algorithm in DRIVE database is 95.39%,the sensitivity is 78.42%,and the F1 value is 0.8213 respectively,which is a certain improvement over the existing algorithms.
作者
薛文渲
刘建霞
XUE Wenxuan;LIU Jianxia(College of Information and Computer,Taiyuan University of Technology,Jinzhong 030600,China)
出处
《电子设计工程》
2021年第20期165-168,共4页
Electronic Design Engineering
基金
2020年度山西省研究生教育创新项目(2020SY521)。
