Thesis understanding: "Cross-Scale Residual Network: A GeneralFramework for Image Super-Resolution, Denoising, and "

译：A general framework for cross-scale residual networks

-- IEEE TRANSACTIONS ON CYBERNETICS -- 2020

---

目录

一、引言

二、网络框架

A、Shallow feature extraction stage

B、Hierarchical feature fusion stage

C、重建阶段

三、Cross-scale residual blocks(CSRB)

四、实验结果

---

一、引言

        一般来说,The purpose of image recovery is from corrupted observationsx = H(Y) + vto restore a clean imagey,其中Y是y的ground-truthHigh quality version,His a degenerate function,v是加性噪声.By adapting to different types of degradation functions,The resulting mathematical model is specific to the image restoration task,如图像超分辨率、Denoising and deblocking.

        The authors hope that the image restoration network can well support the above three tasks.But most existing models can only perform well on one of these tasks.总的来说,All of these tasks have a common feature:Aims to generate visually pleasing high-quality images from low-quality images.So these tasks happen to be strongly correlated,The author thought of designing a common framework to support all tasks,Therefore, a cross-scale residual network is proposed(CSRnet).

二、网络框架

        The network proposed in this paperCSRnet包括三个阶段：Shallow feature extraction stage,Hierarchical feature fusion stage,重建阶段.

They are respectively responsible for extracting shallow image features,利用提出的CSRBRich feature maps in ,as well as adding image details.CSRnet的输入和输出分别为x和y.结构图如下：

A、Shallow feature extraction stage

        Extract shallow features from low-quality input images using two convolutional layers.The first convolutional layer performs feature extraction on the input image,The second convolutional layer performs dimensionality reduction on the features.The shallow feature extraction stage can be expressed as

其中Represents the first convolution operation,卷积核大小为7x7.Using a large convolution kernel can produce a large receptive field.Represents the second convolution operation,卷积核大小为3×3.Connect by jumping,It is further used for residual learning in the reconstruction stage,作为第一个CSRB的输入. 

B、Hierarchical feature fusion stage

       This stage has multiple of the same structureCSRBto learn about layering properties.如果D个CSRB全采用inter-block connection的方式进行堆叠,Then this stage is represented as ：

C、重建阶段

        To further improve information flow and reconstruct image details,This stage contains one skip connection and two convolutional layers,表示为

优化目标为

三、Cross-scale residual blocks(CSRB)

        To determine image features at different scales,作者提出了CSRB作为CSRnet的关键组成部分.CSRBTake three branches using different sizes(即1x、1/2x和1/4x),to support the use of cross-scale features.如图2所示,具体如下：

        图2Boxes with different colored borders in , represent state designs at different scales.带黑色、Boxes with purple and yellow borders represent respectivelys = 0、2和4state at scale.sThe value of represents the scale of downsampling,即1x、1/2和1/4.

        第d个CSRBThe total input is (三个尺度)：

 第d次CSRB中不同尺度s = 0,2,4的输出可以表示为：

其中：

四、实验结果

Below is the denoising result：