作者:Bo Wang1,2, Shuang Qiu2, and Huiguang He1,2,3
目的:Retinal Vessel Segmentation is an essential step for the early diagnosis of eye-related diseases, such as diabetes and hypertension. Segmentation of blood vessels requires both sizeable receptive field and rich spatial information.
方法:Dual Encoding U-Net (DEU-Net), 空间information和上下文information
该结构图outputpatches竟然和input一样。
- Spatial Path,连续四个卷积,用了大stride:7,12,9,6.参考了Large kernel matters-improve semantic segmentation by global convolutional network., CVPR2017。
- Context Path,就是inception block,google提出的
- Attention Skip Module,最简单的attention方式处理
- Feature Fusion Module,这个方式我看到过,不知道为什么叫做feature fusion,其实连结处就是和attention residual for image classification那篇文章一样.
- Multiscale Predict Module,这个模块没看到过,主要是pixel shuffle(参考CVPR2016 Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network)这个操作。
试验结果:略