U-Net: Convolutional Networks for Biomedical Image Segmentation
18 May 2015
Network Architecture
结构对称
contracting path to capture context : repeated application of two 3x3 unpadded convolutions + relu + max pooling, 每次downsampling将channel * 2
symmetric expanding path that enables precise localization
The initial weights from a Gaussian distribution with a standard
deviation of p2=N, where N denotes the number of incoming nodes of one neuron
Loss
加权：类别数量+到边界距离
