Reference

B station video

Scenario

1. Medicine
2. Small target segmentation

U-net model

1. Encode-Decode: (encode the image into a feature first, then decode the feature into an image)
2. Feature fusion, the splicing method is better than direct addition (because the spliced ​​features will be weighted and fused in the next step)
   

Figure 1: U-net:

Next is to improve the model

U-net++ model

Improving ideas (feature fusion and fpn-like multi-branch prediction and multi-loss)

In Figure 1$X^{0,0}$ and $X^{0,4}$ are separated by a long distance, so a new path for feature fusion is added in the middle.


Using the intermediate result prediction is equivalent to doing each step well, and the final result is also very good.

Also the model is easy to prune

U-net+++

The main innovation is in the $X^4$ combines both shallow location information and deep semantic information.