2018-8-7, jingwenlai
# 何为UNet?
基于FPN，修改并扩展，加入了同级连接形成UNet，含有更高层语义信息，可有更大的感受野及更好的精度。
此外，通过overlap-tile strategy使得可以无缝切割任意大的图像。
针对细胞粘连的挑战，引入weighted loss,对边界赋予更大的权重。
## 网络结构
如上图1所示，输入为偶数，需要留意的是the concatenation with the correspondingly cropped feature map from teh contracting path. The cropping is necessary due to the loss of border pixels in every convolution.
## 训练
1. favor large input tiles over a large batch size , batch_size=1 (to minimize the overhead and make maximum use of the GPUmemory)
2. high momentum(0.99) [such that a large number of the previously seen training samples determine the update in the curent optimization step]
3. Energy function is computed by a pixel-wise soft-max over the final feature map combined with the cross entropy loss function. 这一段还需要好好再读下。(权重的设计也是本篇出彩的地方，可以结合代码再细读下)
### Data Augmentation
医学数据样本均较少，除传统的shift/rotation/gray value variations之外，还引入一种deformation augmentation: We generate smooth deformations using random displacement vectors on a coarse 3 by 3 grid. The displacements are sampled from a Gaussian distribution with 10 pixels standard deviation. Per-pixel displacements are then computed using bicubic interpolation. [较独特的data augmentation方法]
## 结果
针对医学图像，取得了很好的结果
原始的caffe实现代码可在https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/下载，
仅需要很少的标注数据及很少的训练时间，即达到了前述效果。
作者也相信在其它任务中，U-Net也会有很好的表现。