Understanding of "dbdnet: a deep boosting strategy for imagedenoising"

translate ：DBDnet： Depth enhancement strategy for image denoising

-- IEEE TRANSACTIONS ON MULTIMEDIA   -- 2021

---

Catalog

One 、 introduction

Two 、 Method

A、 motivation

B、NoN Eliminate modules

C、 Network structure

3、 ... and 、 experimental result

---

One 、 introduction

         In the learning based denoising method , Residual learning is the most commonly used denoising method . This method can generate a noise map from noisy observations , Then we can get a noiseless image , Pictured 1 Shown .

However , The method based on residual learning is difficult to obtain accurate noise map even if using complex networks , That is, the noise graph always contains some noise . The author calls this kind of noise the noise of noise graph (NoN). Intuitively ,NoN The more , The lower the denoising performance , Find an effective way to reduce NoN Influence , It is very important to obtain high-quality images .

         Previous methods based on deep learning rarely consider this problem . In the field of traditional image denoising , Based on boosting Methods , And designed some good boosting Module to solve this problem . This module can iteratively eliminate NON, Extract a clean noise map . Therefore, the author developed boosting Algorithm , And make full use of their advantages to eliminate NoN.

Let's introduce boosting Algorithm ：

        In recent years , People have proposed a variety of methods based on boosting Methods ,“twicing” Technology is a very early research , Its boosting The process is described as follows :

among f It is a de-noising operator , The left side of the equation represents the first n Sub iteration .

be based on “twicing” technology ,Bregman Iteration adopts an iterative regularization method , This method is based on Bregman The concept of distance , Add the residual noise back to the observation signal .Bregman Iterative boosting The process can be written as :

In recent years ,boosting The algorithm is introduced into the field of deep learning , To improve the performance of the network . Here is the author's method .

Two 、 Method

A、 motivation

         The basic problem of image denoising is from noise observation y Restore clean images in x, It can be expressed as :

among v Represents additive noise mapping , Gaussian white noise usually modeled as zero mean . The main goal of the residual learning network is to observe from noise y Generate noise mapping v Approximation of , This process can be expressed as :

among F Represents the algorithm for generating noise residuals , The leftmost side of the equation is the generated noise signature (GNFM).

Besides , Noiseless image x The approximate calculation of is :

The left side of the equation is the final denoising result . From this formula , It has a great impact on the final result , However, due to the capacity limitations of general Networks , Will contain some noise ( namely NoN), That is to say v and It's not equal , Suppose there is a gap u, That is, both meet ：.

        The author thinks that u Affected by two parts , The first is the original image x High frequency information ( That is, boundary information and detail information ), Especially when generating In the process of , original image x Some high-frequency information Will be recognized as noise , Introduced to the in , As a result False identification part of . Besides , Noise observation y Although some pixels of are polluted by noise , But it will still be recognized as clean pixels . therefore , Noise map cannot be completely extracted from noise observation , It will contain some unrecoverable noise information .

         under these circumstances ,u It can be used To express . in other words , The unrecovered noise information add , Subtract the misidentified high-frequency information , You can start from Get a clean noise map v. The process can be expressed as :

therefore , stay boost In the process , You can update :

among and It's network simulation and Generated feature mapping .

Through this operation ,NoN Can gradually reduce .

B、NoN Eliminate modules

         stay Eq.(7) Inspired by , eliminate NoN The process can be decomposed into generating characteristic graphs 、 Generate feature map The process of . So , The author puts forward two different NoN Eliminate the module to generate these two feature maps . Next , The specific implementation of these two non elimination modules will be introduced in detail .

(1)、 modular A： modular A It can be downloaded from Simultaneous generation and , The overall structure is shown in the figure 2(A) Shown .

A convolution block is used to extract the unrecovered noise map And a dense attention block to extract false high-frequency information . Be careful , It's noise mapping v Part of , It's easy to get from Extract it from . however Is the original image x Part of, not v Of , So in order to extract it accurately , The author uses the dense attention block ( Strong ability to capture information ). In order to improve the complexity of dense attention blocks and the ability of information capture , The author introduces two advanced deep learning technologies ( Dense connection and attention mechanism ), Pictured 2(d) Shown .

The whole process can be described as ：

CB It's a convolution block ,DAB Indicates a dense attention block .

(2)、 modular B： The author found that from It's extracted directly from and It's not accurate , Because they will affect each other in the process of being extracted , modular B That's the solution , See the picture for details 2(b) Shown . modular B First generate , Then subtract it as the input of the next module . The whole process is shown as follows ：

C、 Network structure

         After obtaining the above two NoN After eliminating the module , Plug them into the network , Generate DBDnet. The network structure is shown in the figure 3 Shown .

Take the noise observation value y For input , Convolution layer extraction , experienced n individual NoN Eliminate modules , Finally, the final predicted noise map is obtained through a convolution layer . The specific algorithm is as follows ：

among It means the first one n individual NoN Eliminate modules .

The optimization objectives are as follows ：

3、 ... and 、 experimental result

         Here are some experimental results .

grayscale ：

among GFLOPs For computational complexity ,FPS To infer speed .

Real images ：

Code ：

pcl111/DBDNet: Code of "DBDnet: A Deep Boosting Strategy for Image Denoising" (github.com)