【论文笔记】图像修复MPRNet:Multi-Stage Progressive Image Restoration 含代码解析

目录

一、介绍

二、使用方法

1.推理

2.训练

三、MPRNet结构

1.整体结构

2.CAB(Channel Attention Block)

3.Stage1 Encoder

4.Stage2 Encoder

5.Decoder

6.SAM(Supervised Attention Module)

7.ORSNet(Original Resolution Subnetwork)

四、损失函数

1.CharbonnierLoss

2.EdgeLoss

---

一、介绍

        论文地址：https://arxiv.org/pdf/2102.02808.pdf

        代码地址：http://github.com/swz30/MPRNet

        恢复图像任务，需要在空间细节和高级上下文特征之间取得复杂的平衡。于是作者设计了一个多阶段的模型，模型首先使用编解码器架构来学习上下文的特征，然后将它们与保留局部信息的高分辨率分支结合起来。

        打个比方，我要修复一张蛇的图片，编解码器负责提取高级上下文特征，告诉模型要在蛇身上“画”鳞片，而不是羽毛或其他东西；然后高分辨率分支负责细化鳞片的图案。

        MPRNet细节很多，但最主要的创新还是“多阶段”，模型共有三个阶段，前两个阶段是编解码器子网络，用来学习较大感受野的上下文特征，最后一个阶段是高分辨率分支，用于在最终的输出图像中构建所需的纹理。作者给出了Deblurring、Denoising、Deraining三个任务的项目，三个项目的backbone是一样的，只是参数规模有所不同(Deblurring>Denoising>Deraining)，下面我们以最大的Deblurring为例进行介绍。

二、使用方法

        MPRNet项目分为Deblurring、Denoising和Deraining  三个子项目。作者没有用稀奇古怪的库，也没用高级的编程技巧，非常适合拿来研究学习，使用方法也很简单，几句话技能说完。

1.推理

        （1）下载预训练模型：预训练模型分别存在三个子项目的pretrained_models文件夹，下载地址在每个pretrained_models文件夹的 README.md中，需要科学上网，我放在了网盘里：

        链接：https://pan.baidu.com/s/1sxfidMvlU_pIeO5zD1tKZg 提取码：faye 

        （2）准备测试图片：将退化图片放在目录samples/input/中

        （3）执行demo.py

# 执行Deblurring
python demo.py --task  Deblurring

# 执行Denoising
python demo.py --task  Denoising

# 执行Deraining
python demo.py --task  Deraining

        （4）结果放在目录samples/output/中。

2.训练

        （1）根据Dataset文件夹内的README.md文件中的地址下载数据集。

        （2）查看training.yml是否需要修改，主要是最后的数据集地址。

        （3）执行训练

python train.py

三、MPRNet结构

        我将按照官方代码实现来介绍模型结构，一些重要模块的划分可能跟论文有区别，但是整体结构是一样的。

1.整体结构

        MPRNet官方给出的结构图如下：

图1

        这个图总体概括了MPRNet的结构，但是很多细节没有表现出来，通过阅读代码我给出更加详细的模型结构介绍。下面的图中输入统一512x512，我们以Deblurring为例，并且batch_size=1。

        整体结构图如下：

 图2

        图中的三个Input都是原图，整个模型三个阶段，整体流程如下：

        1.1 输入图片采用multi-patch方式分成四份，分成左上、右上、左下、右下；

        1.2 每个patch经过一个3x3的卷积扩充维度，为的是后面能提取更丰富的特征信息；

        1.3 经过CAB(Channel Attention Block)，利用注意力机制提取每个维度上的特征；

        1.4 Encoder，编码三种尺度的图像特征，提取多尺度上下文特征，同时也是提取更深层的语义特征；

        1.5 合并深特征，将四个batch的同尺度特征合并成左右两个尺度，送入Decoder；

        1.6 Decoder，提取合并后的每个尺度的特征；

        1.7 输入图片采用multi-patch方式分成两份，分成左、右；

        1.8 将左右两个batch分别与Stage1 Decoder输出的大尺度特征图送入SAM（Supervised Attention Module），SAM在训练的时候可以利用GT为当前阶段的恢复过程提供有用的控制信号；

        1.9 SAM的输出分成两部分，一部分是第二次输入的原图特征，它将继续下面的流程；一部分用于训练时的Stage1输出，可以利用GT更快更好的让模型收敛。

        2.0 经过Stage2的卷积扩充通道和CAB操作，将Stage1中的Decoder前后的特征送入Stage2的Encoder。

        2.2 经过和Stage1相似的Decoder，也产生两个部分的输出，一部分继续Stage3，一部分输出与GT算损失；

        3.1 Stage3的原图输入不在切分，目的是利用完整的上下文信息恢复图片细节。

        3.2 将原图经过卷积做升维处理；

        3.3 将Stage2中的Decoder前后的特征送入Stage3的ORSNet(Original Resolution Subnetwork)，ORSNet不使用任何降采样操作，并生成空间丰富的高分辨率特征。

        3.4 最后经过一个卷积将维度降为3，输出。

代码实现：

#位置：MPRNet.py
class MPRNet(nn.Module):
    def __init__(self, in_c=3, out_c=3, n_feat=96, scale_unetfeats=48, scale_orsnetfeats=32, num_cab=8, kernel_size=3, reduction=4, bias=False):
        super(MPRNet, self).__init__()

        act=nn.PReLU()
        self.shallow_feat1 = nn.Sequential(conv(in_c, n_feat, kernel_size, bias=bias), CAB(n_feat,kernel_size, reduction, bias=bias, act=act))
        self.shallow_feat2 = nn.Sequential(conv(in_c, n_feat, kernel_size, bias=bias), CAB(n_feat,kernel_size, reduction, bias=bias, act=act))
        self.shallow_feat3 = nn.Sequential(conv(in_c, n_feat, kernel_size, bias=bias), CAB(n_feat,kernel_size, reduction, bias=bias, act=act))

        # Cross Stage Feature Fusion (CSFF)
        self.stage1_encoder = Encoder(n_feat, kernel_size, reduction, act, bias, scale_unetfeats, csff=False)
        self.stage1_decoder = Decoder(n_feat, kernel_size, reduction, act, bias, scale_unetfeats)

        self.stage2_encoder = Encoder(n_feat, kernel_size, reduction, act, bias, scale_unetfeats, csff=True)
        self.stage2_decoder = Decoder(n_feat, kernel_size, reduction, act, bias, scale_unetfeats)

        self.stage3_orsnet = ORSNet(n_feat, scale_orsnetfeats, kernel_size, reduction, act, bias, scale_unetfeats, num_cab)

        self.sam12 = SAM(n_feat, kernel_size=1, bias=bias)
        self.sam23 = SAM(n_feat, kernel_size=1, bias=bias)
        
        self.concat12  = conv(n_feat*2, n_feat, kernel_size, bias=bias)
        self.concat23  = conv(n_feat*2, n_feat+scale_orsnetfeats, kernel_size, bias=bias)
        self.tail     = conv(n_feat+scale_orsnetfeats, out_c, kernel_size, bias=bias)

    def forward(self, x3_img):
        # Original-resolution Image for Stage 3
        H = x3_img.size(2)
        W = x3_img.size(3)

        # Multi-Patch Hierarchy: Split Image into four non-overlapping patches

        # Two Patches for Stage 2
        x2top_img  = x3_img[:,:,0:int(H/2),:]
        x2bot_img  = x3_img[:,:,int(H/2):H,:]

        # Four Patches for Stage 1
        x1ltop_img = x2top_img[:,:,:,0:int(W/2)]
        x1rtop_img = x2top_img[:,:,:,int(W/2):W]
        x1lbot_img = x2bot_img[:,:,:,0:int(W/2)]
        x1rbot_img = x2bot_img[:,:,:,int(W/2):W]

        ##-------------------------------------------
        ##-------------- Stage 1---------------------
        ##-------------------------------------------
        ## Compute Shallow Features
        x1ltop = self.shallow_feat1(x1ltop_img)
        x1rtop = self.shallow_feat1(x1rtop_img)
        x1lbot = self.shallow_feat1(x1lbot_img)
        x1rbot = self.shallow_feat1(x1rbot_img)
        
        ## Process features of all 4 patches with Encoder of Stage 1
        feat1_ltop = self.stage1_encoder(x1ltop)
        feat1_rtop = self.stage1_encoder(x1rtop)
        feat1_lbot = self.stage1_encoder(x1lbot)
        feat1_rbot = self.stage1_encoder(x1rbot)
        
        ## Concat deep features
        feat1_top = [torch.cat((k,v), 3) for k,v in zip(feat1_ltop,feat1_rtop)]
        feat1_bot = [torch.cat((k,v), 3) for k,v in zip(feat1_lbot,feat1_rbot)]
        
        ## Pass features through Decoder of Stage 1
        res1_top = self.stage1_decoder(feat1_top)
        res1_bot = self.stage1_decoder(feat1_bot)

        ## Apply Supervised Attention Module (SAM)
        x2top_samfeats, stage1_img_top = self.sam12(res1_top[0], x2top_img)
        x2bot_samfeats, stage1_img_bot = self.sam12(res1_bot[0], x2bot_img)

        ## Output image at Stage 1
        stage1_img = torch.cat([stage1_img_top, stage1_img_bot],2) 
        ##-------------------------------------------
        ##-------------- Stage 2---------------------
        ##-------------------------------------------
        ## Compute Shallow Features
        x2top  = self.shallow_feat2(x2top_img)
        x2bot  = self.shallow_feat2(x2bot_img)

        ## Concatenate SAM features of Stage 1 with shallow features of Stage 2
        x2top_cat = self.concat12(torch.cat([x2top, x2top_samfeats], 1))
        x2bot_cat = self.concat12(torch.cat([x2bot, x2bot_samfeats], 1))

        ## Process features of both patches with Encoder of Stage 2
        feat2_top = self.stage2_encoder(x2top_cat, feat1_top, res1_top)
        feat2_bot = self.stage2_encoder(x2bot_cat, feat1_bot, res1_bot)

        ## Concat deep features
        feat2 = [torch.cat((k,v), 2) for k,v in zip(feat2_top,feat2_bot)]

        ## Pass features through Decoder of Stage 2
        res2 = self.stage2_decoder(feat2)

        ## Apply SAM
        x3_samfeats, stage2_img = self.sam23(res2[0], x3_img)


        ##-------------------------------------------
        ##-------------- Stage 3---------------------
        ##-------------------------------------------
        ## Compute Shallow Features
        x3     = self.shallow_feat3(x3_img)

        ## Concatenate SAM features of Stage 2 with shallow features of Stage 3
        x3_cat = self.concat23(torch.cat([x3, x3_samfeats], 1))
        
        x3_cat = self.stage3_orsnet(x3_cat, feat2, res2)

        stage3_img = self.tail(x3_cat)

        return [stage3_img+x3_img, stage2_img, stage1_img]

        图中还有一些模块细节没有表现出来，下面我将详细介绍。

2.CAB(Channel Attention Block)

        顾名思义，CAB就是利用注意力机制提取每个通道的特征，输出输入特征图形状不变，结构图如下：

图3 

        可以看到，经过了两个卷积和GAP之后得到了一个概率图（就是那个残差边），在经过两个卷积和Sigmoid之后与概率图相乘，就实现了一个通道注意力机制。

        代码实现：

# 位置MPRNet.py
## Channel Attention Layer
class CALayer(nn.Module):
    def __init__(self, channel, reduction=16, bias=False):
        super(CALayer, self).__init__()
        # global average pooling: feature --> point
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        # feature channel downscale and upscale --> channel weight
        self.conv_du = nn.Sequential(
                nn.Conv2d(channel, channel // reduction, 1, padding=0, bias=bias),
                nn.ReLU(inplace=True),
                nn.Conv2d(channel // reduction, channel, 1, padding=0, bias=bias),
                nn.Sigmoid()
        )

    def forward(self, x):
        y = self.avg_pool(x)
        y = self.conv_du(y)
        return x * y


##########################################################################
## Channel Attention Block (CAB)
class CAB(nn.Module):
    def __init__(self, n_feat, kernel_size, reduction, bias, act):
        super(CAB, self).__init__()
        modules_body = []
        modules_body.append(conv(n_feat, n_feat, kernel_size, bias=bias))
        modules_body.append(act)
        modules_body.append(conv(n_feat, n_feat, kernel_size, bias=bias))

        self.CA = CALayer(n_feat, reduction, bias=bias)
        self.body = nn.Sequential(*modules_body)

    def forward(self, x):
        res = self.body(x)
        res = self.CA(res)
        res += x
        return res

3.Stage1 Encoder

        Stage1和Stage1的Encoder有一些区别，所以分开介绍。Stage1 Encoder有一个输入和三个不同尺度的输出，为的是提取三个尺度的特征并为下面的尺度融合流程做准备；其中有多个CAB结构，可以更好的提取通道特征；下采样通过粗暴的Downsample实现，结构如下：

         图4

        代码实现：

# 位置MPRNet.py
class Encoder(nn.Module):
    def __init__(self, n_feat, kernel_size, reduction, act, bias, scale_unetfeats, csff):
        super(Encoder, self).__init__()

        self.encoder_level1 = [CAB(n_feat,                     kernel_size, reduction, bias=bias, act=act) for _ in range(2)]
        self.encoder_level2 = [CAB(n_feat+scale_unetfeats,     kernel_size, reduction, bias=bias, act=act) for _ in range(2)]
        self.encoder_level3 = [CAB(n_feat+(scale_unetfeats*2), kernel_size, reduction, bias=bias, act=act) for _ in range(2)]

        self.encoder_level1 = nn.Sequential(*self.encoder_level1)
        self.encoder_level2 = nn.Sequential(*self.encoder_level2)
        self.encoder_level3 = nn.Sequential(*self.encoder_level3)

        self.down12  = DownSample(n_feat, scale_unetfeats)
        self.down23  = DownSample(n_feat+scale_unetfeats, scale_unetfeats)

        # Cross Stage Feature Fusion (CSFF)
        if csff:
            self.csff_enc1 = nn.Conv2d(n_feat,                     n_feat,                     kernel_size=1, bias=bias)
            self.csff_enc2 = nn.Conv2d(n_feat+scale_unetfeats,     n_feat+scale_unetfeats,     kernel_size=1, bias=bias)
            self.csff_enc3 = nn.Conv2d(n_feat+(scale_unetfeats*2), n_feat+(scale_unetfeats*2), kernel_size=1, bias=bias)

            self.csff_dec1 = nn.Conv2d(n_feat,                     n_feat,                     kernel_size=1, bias=bias)
            self.csff_dec2 = nn.Conv2d(n_feat+scale_unetfeats,     n_feat+scale_unetfeats,     kernel_size=1, bias=bias)
            self.csff_dec3 = nn.Conv2d(n_feat+(scale_unetfeats*2), n_feat+(scale_unetfeats*2), kernel_size=1, bias=bias)

    def forward(self, x, encoder_outs=None, decoder_outs=None):
        enc1 = self.encoder_level1(x)
        if (encoder_outs is not None) and (decoder_outs is not None):
            enc1 = enc1 + self.csff_enc1(encoder_outs[0]) + self.csff_dec1(decoder_outs[0])

        x = self.down12(enc1)

        enc2 = self.encoder_level2(x)
        if (encoder_outs is not None) and (decoder_outs is not None):
            enc2 = enc2 + self.csff_enc2(encoder_outs[1]) + self.csff_dec2(decoder_outs[1])

        x = self.down23(enc2)

        enc3 = self.encoder_level3(x)
        if (encoder_outs is not None) and (decoder_outs is not None):
            enc3 = enc3 + self.csff_enc3(encoder_outs[2]) + self.csff_dec3(decoder_outs[2])
        
        return [enc1, enc2, enc3]

4.Stage2 Encoder

        Stage2 Encoder输入为三个，分别为上一层的输出和Stage1中的Decoder前后的特征。主流程（也就是左面竖着的那一列）和Stage1 Encoder是一样的。增加的两个输入，每个输入又分为三个尺度，每个尺度经过一个卷积层，然后相同尺度的特征图做特征融合，输出，结构如下：

图5

5.Decoder

        两个阶段的Decoder结构是一样的，所以放在一起说，有三个不用尺度的输入；通过CAB提取特征；小尺度特征通过上采样变大，通过卷积使通道变小；小尺度的特征图shape最终变成跟大尺度一样，通过残差边实现特征融合，结构如下：

 图6

        代码实现：

# 位置：MPRNet.py
class Decoder(nn.Module):
    def __init__(self, n_feat, kernel_size, reduction, act, bias, scale_unetfeats):
        super(Decoder, self).__init__()

        self.decoder_level1 = [CAB(n_feat,                     kernel_size, reduction, bias=bias, act=act) for _ in range(2)]
        self.decoder_level2 = [CAB(n_feat+scale_unetfeats,     kernel_size, reduction, bias=bias, act=act) for _ in range(2)]
        self.decoder_level3 = [CAB(n_feat+(scale_unetfeats*2), kernel_size, reduction, bias=bias, act=act) for _ in range(2)]

        self.decoder_level1 = nn.Sequential(*self.decoder_level1)
        self.decoder_level2 = nn.Sequential(*self.decoder_level2)
        self.decoder_level3 = nn.Sequential(*self.decoder_level3)

        self.skip_attn1 = CAB(n_feat,                 kernel_size, reduction, bias=bias, act=act)
        self.skip_attn2 = CAB(n_feat+scale_unetfeats, kernel_size, reduction, bias=bias, act=act)

        self.up21  = SkipUpSample(n_feat, scale_unetfeats)
        self.up32  = SkipUpSample(n_feat+scale_unetfeats, scale_unetfeats)

    def forward(self, outs):
        enc1, enc2, enc3 = outs
        dec3 = self.decoder_level3(enc3)

        x = self.up32(dec3, self.skip_attn2(enc2))
        dec2 = self.decoder_level2(x)

        x = self.up21(dec2, self.skip_attn1(enc1))
        dec1 = self.decoder_level1(x)

        return [dec1,dec2,dec3]

6.SAM(Supervised Attention Module)

        SAM出现在两个阶段间，有两个输入，将上一层特征和原图作为输入，提升了特征提取的性能，，SAM作为有监督的注意模块，使用注意力图强力筛选了跨阶段间的有用特征。有两个输出，一个是经过了注意力机制的特征图，为下面的流程提供特征；一个是3通道的图片特征，为了训练阶段输出，结构如下：

 图7

        代码位置：

# 位置MPRNet.py
## Supervised Attention Module
class SAM(nn.Module):
    def __init__(self, n_feat, kernel_size, bias):
        super(SAM, self).__init__()
        self.conv1 = conv(n_feat, n_feat, kernel_size, bias=bias)
        self.conv2 = conv(n_feat, 3, kernel_size, bias=bias)
        self.conv3 = conv(3, n_feat, kernel_size, bias=bias)

    def forward(self, x, x_img):
        x1 = self.conv1(x)
        img = self.conv2(x) + x_img
        x2 = torch.sigmoid(self.conv3(img))
        x1 = x1*x2
        x1 = x1+x
        return x1, img

7.ORSNet(Original Resolution Subnetwork)

        为了保留输入图像的细节，模型在最后一阶段引入了原始分辨率的子网(ORSNet:Original Resolution Subnetwork)。ORSNet不使用任何降采样操作，并生成空间丰富的高分辨率特征。它由多个原始分辨率块（BRB）组成，是模型的最后阶段，结构如下：

 图8

        可以看到，输入为三个，分别为上一层的输出和Stage2中的Decoder前后的特征。后两个输入，每个输入又分为三个尺度，三个尺度的通道数都先变成96，然后在变成128；小尺度的size都变成和大尺度一样，最后做特征融合融合前会经过ORB(Original Resolution Block)模块。

        ORB由一连串的CAB组成，还有一个大的残差边，结构如下：

 图9

        代码实现：

# 位置MPRNet.py
## Original Resolution Block (ORB)
class ORB(nn.Module):
    def __init__(self, n_feat, kernel_size, reduction, act, bias, num_cab):
        super(ORB, self).__init__()
        modules_body = []
        modules_body = [CAB(n_feat, kernel_size, reduction, bias=bias, act=act) for _ in range(num_cab)]
        modules_body.append(conv(n_feat, n_feat, kernel_size))
        self.body = nn.Sequential(*modules_body)

    def forward(self, x):
        res = self.body(x)
        res += x
        return res

##########################################################################
class ORSNet(nn.Module):
    def __init__(self, n_feat, scale_orsnetfeats, kernel_size, reduction, act, bias, scale_unetfeats, num_cab):
        super(ORSNet, self).__init__()

        self.orb1 = ORB(n_feat+scale_orsnetfeats, kernel_size, reduction, act, bias, num_cab)
        self.orb2 = ORB(n_feat+scale_orsnetfeats, kernel_size, reduction, act, bias, num_cab)
        self.orb3 = ORB(n_feat+scale_orsnetfeats, kernel_size, reduction, act, bias, num_cab)

        self.up_enc1 = UpSample(n_feat, scale_unetfeats)
        self.up_dec1 = UpSample(n_feat, scale_unetfeats)

        self.up_enc2 = nn.Sequential(UpSample(n_feat+scale_unetfeats, scale_unetfeats), UpSample(n_feat, scale_unetfeats))
        self.up_dec2 = nn.Sequential(UpSample(n_feat+scale_unetfeats, scale_unetfeats), UpSample(n_feat, scale_unetfeats))

        self.conv_enc1 = nn.Conv2d(n_feat, n_feat+scale_orsnetfeats, kernel_size=1, bias=bias)
        self.conv_enc2 = nn.Conv2d(n_feat, n_feat+scale_orsnetfeats, kernel_size=1, bias=bias)
        self.conv_enc3 = nn.Conv2d(n_feat, n_feat+scale_orsnetfeats, kernel_size=1, bias=bias)

        self.conv_dec1 = nn.Conv2d(n_feat, n_feat+scale_orsnetfeats, kernel_size=1, bias=bias)
        self.conv_dec2 = nn.Conv2d(n_feat, n_feat+scale_orsnetfeats, kernel_size=1, bias=bias)
        self.conv_dec3 = nn.Conv2d(n_feat, n_feat+scale_orsnetfeats, kernel_size=1, bias=bias)

    def forward(self, x, encoder_outs, decoder_outs):
        x = self.orb1(x)
        x = x + self.conv_enc1(encoder_outs[0]) + self.conv_dec1(decoder_outs[0])

        x = self.orb2(x)
        x = x + self.conv_enc2(self.up_enc1(encoder_outs[1])) + self.conv_dec2(self.up_dec1(decoder_outs[1]))

        x = self.orb3(x)
        x = x + self.conv_enc3(self.up_enc2(encoder_outs[2])) + self.conv_dec3(self.up_dec2(decoder_outs[2]))

        return x

四、损失函数

        MPRNet主要使用了两个损失函数CharbonnierLoss和EdgeLoss，公式如下：

         其中累加是因为训练的时候三个阶段都有输出，都需要个GT计算损失（如图2的三个output）；该模型不是直接预测恢复的图像，而是预测残差图像，添加退化的输入图像得到：  

        Deblurring和Deraining两个任务CharbonnierLoss和EdgeLoss做了加权求和，比例1:0.05；只使用了CharbonnierLoss，我感觉是因为这里使用的噪声是某种分布（入高斯分布、泊松分布）的噪声，不会引起剧烈的边缘差异，所以Denoising没有使用EdgeLoss。

        下面简单介绍一下两种损失。

1.CharbonnierLoss

        公式如下：

        CharbonnierLoss在零点附近由于常数的存在，梯度不会变成零，避免梯度消失。函数曲线近似L1损失，相比L2损失而言，对异常值不敏感，避免过分放大误差。

        代码实现：

# 位置losses.py
class CharbonnierLoss(nn.Module):
    """Charbonnier Loss (L1)"""

    def __init__(self, eps=1e-3):
        super(CharbonnierLoss, self).__init__()
        self.eps = eps

    def forward(self, x, y):
        diff = x - y
        # loss = torch.sum(torch.sqrt(diff * diff + self.eps))
        loss = torch.mean(torch.sqrt((diff * diff) + (self.eps*self.eps)))
        return loss

2.EdgeLoss

        L1或者L2损失注重的是全局，没有很好地考虑一些显著特征的影响， 而显著的结构和纹理信息与人的主观感知效果高度相关，是不能忽视的。

        边缘损失主要考虑纹理部分的差异，可以很好地考虑高频的纹理结构信息， 提高生成图像的细节表现，公示如下：

        其中表示Laplacian边缘检测中的核函数，表示对做边缘检测，公式中其他部分和CharbonnierLoss类似。

        代码实现：

# 位置losses.py

class EdgeLoss(nn.Module):
    def __init__(self):
        super(EdgeLoss, self).__init__()
        k = torch.Tensor([[.05, .25, .4, .25, .05]])
        self.kernel = torch.matmul(k.t(),k).unsqueeze(0).repeat(3,1,1,1)
        if torch.cuda.is_available():
            self.kernel = self.kernel.cuda()
        self.loss = CharbonnierLoss()

    def conv_gauss(self, img):
        n_channels, _, kw, kh = self.kernel.shape
        img = F.pad(img, (kw//2, kh//2, kw//2, kh//2), mode='replicate')
        return F.conv2d(img, self.kernel, groups=n_channels)

    def laplacian_kernel(self, current):
        filtered    = self.conv_gauss(current)    # filter
        down        = filtered[:,:,::2,::2]               # downsample
        new_filter  = torch.zeros_like(filtered)
        new_filter[:,:,::2,::2] = down*4                  # upsample
        filtered    = self.conv_gauss(new_filter) # filter
        diff = current - filtered
        return diff

    def forward(self, x, y):
        loss = self.loss(self.laplacian_kernel(x), self.laplacian_kernel(y))
        return loss

        MPRNet的主要的内容就介绍到这，主要是backbone的创新，其他部分中规中矩，关注不迷路。