• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 图形与图像 • 上一篇    下一篇

基于YOLO算法的眼底图像视盘定位方法

蒋芸,彭婷婷,谭宁,侯金泉   

  1. (西北师范大学计算机科学与工程学院,甘肃 兰州 730070)
  • 收稿日期:2018-08-03 修回日期:2018-10-17 出版日期:2019-09-25 发布日期:2019-09-25
  • 基金资助:

    国家自然科学基金(61163036);甘肃省科技计划资助自然科学基金(1606RJZA047);2012年度甘肃省高校基本科研业务费专项资金项目;甘肃省高校研究生导师项目(1201-16);西北师范大学第三期知识与创新工程科研骨干项目(nwnu-kjcxgc-03-67)

An optic disc positioning method in
fundus images based on YOLO

JIANG Yun,PENG Ting-ting,TAN Ning,HOU Jin-quan   

  1. (College of Computer Science and Engineering,Northwest Normal University,Lanzhou 730070,China)
  • Received:2018-08-03 Revised:2018-10-17 Online:2019-09-25 Published:2019-09-25

摘要:

视盘的各个参数是衡量眼底健康状况和病灶的重要指标,视盘的检测和定位对于观察视盘的形态尤为重要。在以往的视盘定位研究中,主要根据视盘的形状、亮度、眼底血管的走向等特征使用图像处理的方法对眼底图像中视盘进行定位。由于人为因素影响较大,特征提取时间较长,且视盘定位效率低,因此提出一种基于YOLO算法的眼底图像视盘定位方法。利用YOLO算法将眼底图像划分为N×N的格子,每个格子负责检测视盘中心点是否落入该格子中,通过多尺度的方式和残差层融合低级特征对视盘进行定位,得到不同大小的边界框,最后通过非极大抑制的方式筛选出得分最高的边界框。通过在3个公开的眼底图像数据集(DRIVE、DRISHTI-GS1和MESSIDOR)上,对所提出的视盘定位方法进行测试,定位准确率均为100%,实验同时定位出视盘的中心点坐标,与标准中心点的平均欧氏距离分别为2236 px、2.52 px、21.42 px,验证了该方法的准确性和通用性。

 

关键词: 视盘, YOLO算法, 目标检测, 深度学习, 卷积神经网络

Abstract:

The parameters of the optic disc are important indicators for measuring the health status and lesions of the fundus. The detection and localization of the optic disc is especially important for observing the shape of the optic disc. Research on optic disc positioning in the past mainly depends on the shape and brightness of the optic disc, and the direction of the fundus blood vessels, and image-processing methods are used to locate the optic disc in fundus images. Due to the influence of human factors, the feature extraction time is long and the optic disc positioning efficiency is low. We propose a method of locating optic disc of the fundus image based on the YOLO algorithm. The YOLO algorithm is used to divide the fundus image into N×N grids, and each grid is responsible for detecting whether the center point of the disc falls into the grid. The multi-scale method and the residual layer are fused with low-level features to locate the disc, and bounding boxes of different sizes are obtained. The bounding box with the highest score is finally selected through non-maximum suppression. We test the proposed localization method on three open databases of fundus images (DRIVE、DRISHTI-GS1 and MESSIDOR). The positioning accuracy is 100%, and the center point coordinates of the optic disc are located in the experiment. The average Euclidean distances to the center points are 22.36 px, 2.52 px, 21.42 px, respectively, which verifies the accuracy and versatility of the method.
 

Key words: optic disc, YOLO algorithm, object detection, deep learning, convolutional neural network