摘要
针对现有的算法对视盘边缘分割精度不高和视盘周围存在大量噪音等难点,提出一种融合空洞卷积与注意门的U型卷积神经网络视盘分割算法。在预处理阶段,利用图像RGB通道的线性组合提取各通道颜色特征信息,利用形态学滤波技术强化视盘边缘信息。在分割阶段,利用多尺度空洞卷积提高感受野,通过注意门提升视盘权重信息,由SoftMax激活函数分割视盘与背景信息。在DRIONS-DB眼底图像数据集上进行的仿真结果表明,视盘分割准确率、精确率和重合率分别达到99.83%,94.84%和97.35%。
Aiming at the difficulties of utilizing existing algorithms such as low precision of optic disk edge segmentation and lots of noise around the optic disk,a U-shaped convolution neural network optic disk segmentation algorithm based on dilated convolution and attention gate was proposed.In the preprocessing stage,the color feature information of each channel was extracted by the linear combination of RGB channels,and multi-scale morphological filtering was used to enhance the edge information of the optic disk.In the segmentation stage,multi-scale dilated convolution was used to improve the receptive field,and attention gates were used to enhance the optic disk weight information.The disc was segmented using the SoftMax activation function.The simulation experiments on DRIONS-DB fundus image data set show that the accuracy,precision and coincidence rates of optic disc segmentation are 99.83%,94.84%and 97.35%respectively.
作者
梁礼明
盛校棋
熊文
郭凯
LIANG Li-ming;SHENG Xiao-qi;XIONG Wen;GUO Kai(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China)
出处
《计算机工程与设计》
北大核心
2020年第3期808-814,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(51365017、61463018)
江西省自然科学基金项目(20132BAB203020)
江西省教育厅科学技术研究重点基金项目(GJJ170491)。
关键词
空洞卷积
注意门
卷积神经网路
视盘分割
形态学滤波
dilated convolution
attention gates
convolutional neural network
optic disk
morphological filtering
