Know that the account with the same name is published synchronously

One 、 framework

Full title of the paper ：Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections

In fact, that is conv and deconv, Plus symmetrical skip connection. The author says there will be one on every two floors skip connection.

The author means to say ,conv The function of is to feature extraction, Keep the main components of the objects in the diagram , At the same time eliminate corruption.deconv Is that recover the details of image contents.

All layers are followed by ReLU,feature maps Yes 64 layer （ The author says more can improve , But not much ）,filter size yes 3×3（ This point refers to VGG）. Input image Of size It can be random , The author of the experiment on the input channel The settings are 1.️ Be careful not to pooling and unpooling, The author's reason is that this will discard essential image detail.

Look at another picture ：

This is it. paper Description of the model architecture in , There's nothing to say , Next, let's talk directly about the idea of this article .

Two 、main contribution

1,skip connection

The author says that this connection mainly does two things , One thing is to help gradient back propagation , It's easy to understand .

Another thing is to help transfer the details of the image to the later layers （pass image details to the top layers）.

The advantage of doing so is , On the one hand, it can improve performance, On the other hand, it can network Become deeper .

2, The first method

The author says this is the first to use single model Make a difference noise level And realize the denoising task good accuracy Methods .（ But notice that it is 2016.9.1,DnCNN yes 2016.8.13, The latter can also achieve different noise level Denoising of , however level Must be in preset Of range Inside ）.

3,performance leading

More recently （2016.9.1） De-noising and super fractional SOTA All right .

At the same time, pay attention to the author's understanding of DNN methods There is a description ：purely data driven; no assumption about noise distribution.

3、 ... and 、 Some discussions

1, If not skip connection：

If not , So for shallow networks ,deconv can recover detail; For deep networks, you can't . So this connection is for recover detail And it helps to make the network deeper .

2, and highway nets,resnet Different ：

The network pass information of the conv feature maps to the corresponding deconv layers.

3,residual learning

What network studies is residuals , No noisy image To noise-free image Mapping .

Four 、 Training & Model function & Some comparisons

1,100 individual epochs after , from loss Look up ,30 Layer belt connection ＜20 Layer belt connection ＜10 Layer without connection ＜20 Layer without connection ＜30 Layer without connection . The connection here refers to skip connection, It can be seen that , If you don't bring it , If there are more layers, there will be problems . therefore Proved skip connection The necessity of .

2, and resnet It is also compared , identical block size when ,PSNR On RED-Net Better than resnet. The author advocates their skip connection Yes element-wise correspondence, This is right pixel-wise prediction problems Very important .
（block size Refers to the span of the connection （the span of the connections））

3,loss Euclidean distance is used , Specifically, minimize the following ：

4, There is one trick It's the last layer with a smaller learning rate , The author doesn't think it's necessary here .

5,* How to generate training sets ：
The customary ：

• gray-scale image for denoising
• luminance channel for super-resolution

use 300 Zhang BSD Of images,50×50 Of patch As ground-truth：

• Generate noisy image： Add additive Gaussian noise ;
• Generate low resolution maps ： First down sampling , And then sampling to original size

The above two kinds of data are for denoising and super-resolution tasks respectively , This model uses these two to train , Be able to handle two tasks at the same time .（️ Be careful ： In the training dataset , Generate noisy image Using a different noise level, Different methods are used to generate low resolution images scaling parameter）

5、 ... and 、 Summary

This article is very simple , Sum up , The model has two parts of work , Namely extracting primary image content and recovering details. The point is to use skip connections, The author's description is ：

which helps on recovering clean images and tackles the optimization difficulty caused by gradient vanishing, and thus obtains performance gains when the network goes deeper

That is to help solve the optimization difficulties caused by the disappearance of gradient , So the network can become deeper, thus gain some performance The promotion of . Then it is better than that at that time in denoising and super division SOTA.