《BRDNet+ADNet》阅读笔记
一、论文
《Image denoising using deep CNN with batch renormalization》
深度卷积神经网络（CNN）在图像去噪领域引起了极大的关注。 但是，有两个缺点：（1）很难训练更深的CNN来进行去噪任务；（2）大多数更深的CNN都存在性能饱和的问题。 在本文中，我们报告了一种称为批重归一化去噪网络（BRDNet）的新型网络的设计。 具体来说，我们将两个网络合并以增加网络的宽度，从而获得更多功能。 因为批处理重归一化已融合到BRDNet中，所以我们可以解决内部协变量偏移和小型迷你批处理问题。 还采用整体学习的残差学习方式来促进网络培训。 利用扩张卷积来提取更多信息以进行去噪任务。 大量的实验结果表明，BRDNet优于最新的图像降噪方法。
《Attention-guided CNN for Image Denoising》
深度卷积神经网络（CNN）在低级计算机视觉中引起了相当大的兴趣。 研究通常致力于通过非常深的CNN来提高性能。 但是，随着深度的增加，浅层对深层的影响会减弱。受这一事实的启发，我们提出了一种注意力导向的去噪卷积神经网络（ADNet），主要包括稀疏块（SB），特征增强块（FEB），注意块（AB）和重构块（RB） 图像降噪。 具体来说，SB通过使用膨胀和普通卷积来去除噪声，从而在性能和效率之间进行权衡。 FEB通过很长的路径集成了全局和局部特征信息，以增强去噪模型的表达能力。 AB用于精细提取隐藏在复杂背景中的噪声信息，这对于复杂的噪点图像（尤其是真实的噪点图像）和绑定去噪非常有效。 而且，FEB与AB集成在一起，可以提高效率并降低训练降噪模型的复杂度。 最终，RB旨在通过获得的噪声映射和给定的噪声图像来构造清晰图像。 此外，全面的实验表明，拟议的ADNet在三个任务中表现出色（即 定量和定性评估方面的合成和真实噪点图像，以及盲降噪）。
二、模型结构
1、BRDNet
2、ADNet
三、代码
BRDNet代码下载：https://github.com/hellloxiaotian/BRDNet
ADNet代码下载：https://github.com/hellloxiaotian/ADNet
import torch
import torch.nn as nn
class Conv_BN_Relu_first(nn.Module):
def __init__(self,in_channels,out_channels,kernel_size,padding,groups,bias):
super(Conv_BN_Relu_first,self).__init__()
kernel_size = 3
padding = 1
features = 64
groups =1
self.conv = nn.Conv2d(in_channels=channels, out_channels=features, kernel_size=kernel_size, padding=padding,groups=groups, bias=False)
self.bn = nn.BatchNorm2d(features)
self.relu = nn.ReLU(inplace=True)
def forward(self,x):
return self.relu(self.bn(self.conv(x)))
class Conv_BN_Relu_other(nn.Module):
def __init__(self,in_channels,out_channels,kernel_size,padding,groups,bias):
super(Conv_BN_Relu_other,self).__init__()
kernel_size = 3
padding = 1
features = out_channels
groups =1
self.conv = nn.Conv2d(in_channels=in_channels, out_channels=features, kernel_size=kernel_size, padding=padding,groups=groups, bias=False)
self.bn = nn.BatchNorm2d(features)
self.relu = nn.ReLU(inplace=True)
def forward(self,x):
return self.relu(self.bn(self.conv(x)))
class Conv(nn.Module):
def __init__(self,in_channels,out_channels,kernel_size,padding,groups,bais):
super(Conv,self).__init__()
kernel_size = 3
padding = 1
features = 1
groups =1
self.conv = nn.Conv2d(in_channels=channels, out_channels=features, kernel_size=kernel_size, padding=padding,groups=groups, bias=False)
def forward(self,x):
return self.conv(x)
class Self_Attn(nn.Module):
def __init__(self,in_dim):
super(Self_Attn,self).__init__()
self.chanel_in = in_dim
self.query_conv = nn.Conv2d(in_channels=in_dim,out_channels=in_dim//8,kernel_size=1)
self.key_conv = nn.Conv2d(in_channels=in_dim,out_channels=in_dim//8,kernel_size=1)
self.value_conv = nn.Conv2d(in_channels=in_dim,out_channels=in_dim,kernel_size=1)
self.gamma=nn.Parameter(torch.zeros(1))
self.softmax=nn.Softmax(dim=-1)
def forward(self,x):
m_batchsize, C, width,height = x.size()
proj_query = self.query_conv(x).view(m_batchsize,-1,width*height).permute(0,2,1)
proj_key = self.key_conv(x).view(m_batchsize,-1,width*height)
print proj_query.size()
print proj_key.size()
print '5'
energy = torch.bmm(proj_query,proj_key)
print '6'
#print energy.size()
attention = self.softmax(energy)
proj_value = self.value_conv(x).view(m_batchsize,-1,width*height)
print '1'
out = torch.bmm(proj_value,attention.permute(0,2,1))
print '2'
out = out.view(m_batchsize,C,width,height)
out = self.gamma*out + x
return out, attention
class ADNet(nn.Module):
def __init__(self, channels, num_of_layers=15):
super(ADNet, self).__init__()
kernel_size = 3
padding = 1
features = 64
groups =1
layers = []
kernel_size1 = 1
'''
#self.gamma = nn.Parameter(torch.zeros(1))
'''
self.conv1_1 = nn.Sequential(nn.Conv2d(in_channels=channels,out_channels=features,kernel_size=kernel_size,padding=padding,groups=groups,bias=False),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_2 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=2,groups=groups,bias=False,dilation=2),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_3 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=1,groups=groups,bias=False),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_4 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=1,groups=groups,bias=False),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_5 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=2,groups=groups,bias=False,dilation=2),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_6 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=1,groups=groups,bias=False),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_7 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=padding,groups=groups,bias=False),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_8 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=1,groups=groups,bias=False),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_9 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=2,groups=groups,bias=False,dilation=2),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_10 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=1,groups=groups,bias=False),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_11 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=1,groups=groups,bias=False),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_12 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=2,groups=groups,bias=False,dilation=2),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_13 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=padding,groups=groups,bias=False),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_14 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=padding,groups=groups,bias=False),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_15 = nn.Sequential(nn.Conv2d(in_channels=features,out_channels=features,kernel_size=kernel_size,padding=1,groups=groups,bias=False),nn.BatchNorm2d(features),nn.ReLU(inplace=True))
self.conv1_16 = nn.Conv2d(in_channels=features,out_channels=1,kernel_size=kernel_size,padding=1,groups=groups,bias=False)
self.conv3 = nn.Conv2d(in_channels=2,out_channels=1,kernel_size=1,stride=1,padding=0,groups=1,bias=True)
self.ReLU = nn.ReLU(inplace=True)
self.Tanh= nn.Tanh()
self.sigmoid = nn.Sigmoid()
for m in self.modules():
if isinstance(m, nn.Conv2d):
# n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, (2 / (9.0 * 64)) ** 0.5)
if isinstance(m, nn.BatchNorm2d):
m.weight.data.normal_(0, (2 / (9.0 * 64)) ** 0.5)
clip_b = 0.025
w = m.weight.data.shape[0]
for j in range(w):
if m.weight.data[j] >= 0 and m.weight.data[j] < clip_b:
m.weight.data[j] = clip_b
elif m.weight.data[j] > -clip_b and m.weight.data[j] < 0:
m.weight.data[j] = -clip_b
m.running_var.fill_(0.01)
def _make_layers(self, block,features, kernel_size, num_of_layers, padding=1, groups=1, bias=False):
layers = []
for _ in range(num_of_layers):
layers.append(block(in_channels=features, out_channels=features, kernel_size=kernel_size, padding=padding, groups=groups, bias=bias))
return nn.Sequential(*layers)
def forward(self, x):
input = x
x1 = self.conv1_1(x)
x1 = self.conv1_2(x1)
x1 = self.conv1_3(x1)
x1 = self.conv1_4(x1)
x1 = self.conv1_5(x1)
x1 = self.conv1_6(x1)
x1 = self.conv1_7(x1)
x1t = self.conv1_8(x1)
x1 = self.conv1_9(x1t)
x1 = self.conv1_10(x1)
x1 = self.conv1_11(x1)
x1 = self.conv1_12(x1)
x1 = self.conv1_13(x1)
x1 = self.conv1_14(x1)
x1 = self.conv1_15(x1)
x1 = self.conv1_16(x1)
out = torch.cat([x,x1],1)
out= self.Tanh(out)
out = self.conv3(out)
out = out*x1
out2 = x - out
return out2四、学习资料
效果好速度还好的图像去噪：BRDNet
首个硬件资源受限下数据不均匀的图像去噪网络：BRDNet
哈工大与北大提出注意力引导的图像去噪：ADNet