Know that the account with the same name is released synchronously

This is a unpaired Methods , That is, there are many noisy images, There's a lot of clean images, But a couple noisy-clean images The content is not paired .

Pre knowledge

Need to know CycleGAN Knowledge , You can go to b Stand and watch Li Hongyi's class ,CycleGAN There was only 20 minute , It's very clear .

Model architecture

Suppose you already understand CycleGAN Principle , Then you can pay attention to ISCL The core vocabulary of ：self-cooperative. The model architecture is as follows ：

In the picture above F and G yes CycleGAN It's already there generator,H It's a network that extracts noise , It's a new addition to the author .x representative noisy image,y representative clean image,n representative noise.

Four red lines , The top two are cycle consistency, yes CycleGAN Something that already exists , The next two are called bypass consistency.bypass consistency This is the most important innovation of this work , Then we will recognize its importance through data .

besides , There are other details ：

In the lower right corner of the left figure above $D_y$ And the one in the lower right corner of the figure $D_x$, And the new neural network H dependent discriminator, The result loss, The author calls it boosting loss.

Above is the pseudo noise label , The meaning is self-evident .

The specific networks used are as follows , Just take a simple look ：

because F and H Can provide denoising results , So the two should be used together ：

The above is the description of the model architecture , It's a little bit easier , The formula will not be copied and pasted .

experiment

• LDCT(Mayo)

The main highlights are the following table ：

In the table above (A) Means original CycleGAN,(A)~(E) Represents the author's model after gradually adding a pile of things . It can be seen that , After adding , There are obvious faults in the data , And it means bypass consistency, This also proves that it is indeed the most important innovation of this work .

effect ：

• Synthetic noisy EM denoising

effect ：

• Real noisy EM denoising

effect ：

Summary

There is... In the title self-cooperative, In fact, that is CycleGAN Of generator F And the author's new addition ResNet H Mutual cooperation , So as to improve the effect .