1. The overall network structure of UNet++
Unet++ mainly explores the problem of how big the encoder and decoder need to be. Based on this, a network that integrates Unet structures of different sizes is proposed. The innovation point is to integrate Uent structures of different sizes into a network.
2. The benefits of the structure
(1) No matter which deep feature is effective, use it, let the network learn the importance of different deep features by itself.
(2) An encoder is shared, that is to say, there is no need to train Unet, but only train one encoder, and its different levels of features are restored by different decoder paths.
3. Existing problems
Cannot train because backpropagation is not possible. There are two solutions:
(1) Use short connections.
(2) Use deep supervision on the basis of short connections.
Adding a 1*1 convolution kernel after X0,1, X0,2, X0,3, and X0,4 in the figure is equivalent to supervising the output of Unet of each branch. This solves the problem of not being able to train.
4. Advantages:
It can be easier to prune, because there is also a separate supervised training in front, and the prune can be quickly completed according to the speed requirement.
Why can it be pruned?
In the test phase, since the input image can only be propagated forward, throwing away this part has no effect on the output of the money face at all. In the training phase, because there are both forward and backward propagation, the cut part It will help other parts to update the weight. That is, during testing, the cut-off part has no effect on the remaining structure, and during training, the cut-off part has an impact on the remaining part.
(5) Benefits of pruning
The advantage of pruning is: if the effect of low-scale Uent++ training is equivalent to that of high-scale Unet++ training, only low-scale training will be performed, which will reduce parameters during training and speed up training.