Article directory
1. Summary of FCN
(1) The first end-to-end fully convolutional neural network for pixel-level prediction
(2) Full convolution: replace all the fully connected layers in the classification classic network with convolutional layers
(3) combine the shallow positional features with The deep semantic features are combined, but the combination method is simple addition, which is different from the channel splicing method of U-net.
The FCN network structure is as follows:
2. U-net summary
(1) The fully symmetrical U-shaped structure makes the fusion of front and rear features more thorough, making high-resolution information and low-resolution information increase in the target image.
(2) The way of fusion during skip connection is splicing according to the channel
The network structure of U-net is as follows:
3. Summary of U-net++
The main idea of U-net++ is to add more paths between U-net's encoder and decoder. The left side of U-net++ is downsampling. In the process of downsampling, each downsampled node will be merged with the previous node, and it will be cycled in turn. The goal of U-net++ is to improve segmentation accuracy by adding dense blocks and convolutional layers between the encoder and decoder.
