31
2017 IOS Press
Automatic Identification of Glaucoma Using Deep Learning Methods
Method : 分類
Dataset:HRF , RIM-ONE r1,RIM-ONE r2,RIM-ONE r3
Architecture: ROI + GoogLeNet
Results: AC HRF 90%, RIM-ONE r1 94.2% ,RIM-ONE r2 86.2% ,RIM-ONE r3 86.4% 總 87.6%
眼底圖像缺陷:very subtle lesions, poor image quality, and illumination problems
The purpose of our work : to perform this detection even in images with
- low contrast
- high amount of noise
- low resolution
- ROI barely visible
Common practice :
- post-processing
- remove blood vessels
Aim
- robust network
Method
- find ROI
GoogLeNet 將圖像分類爲 ROI和背景
開發了一種利用滑動窗口掃描 以搜索該區域的算法
- data augmentation
adding deformations and noise
- clasification
GoogLeNet
Disscussion
- 先提取 ROI ,數據增強 ,分類 (沒什麼新思想)
- 擴大數據
- 圖像預處理