由于需要增添对比实验，于是复现了BRDNet去噪网络，并且将其修改为一维信号去噪网络。
- 网络架构
BRDNet网络架构，由两个不同的网络组成：上层网络和下层网络。上层网络仅由RL和BRN组成。较低层网络包括BRN，RL和膨胀卷积。
众所周知，接收域更大，所设计的网络将具有更高的计算成本。因此，我们选择一个网络（下层网络）来使用膨胀卷积。考虑到性能和效率之间的平衡，较低网络的2–8和10–15层使用膨胀卷积来捕获更多上下文信息。第一，第十六层使用BRN对数据进行归一化，这使两个子网的输出保持相同的分布。此外，BRN对于小批量任务非常有用，这对于低配置的硬件平台（例如GTX960和GTX970）非常有用。下一个，RL技术被融合到两个通道的网络中以提高性能。
具体代码如下：
import torch
import torch.nn as nn
from dncnns.model.torchsummary import summary
class BRDNet(nn.Module):
def __init__(self, channels, num_of_layers=15):
super(BRDNet, self).__init__()
kernel_size = 3
padding = 1
features = 64
groups = 1
L = []
layers = []
layers.append(nn.Conv1d(in_channels=channels, out_channels=features, kernel_size=kernel_size, padding=padding,
bias=False))
layers.append(nn.BatchNorm1d(features))
layers.append(nn.ReLU(inplace=True))
# layers.append(B.ResBlock(features, features, kernel_size=kernel_size, stride=1, padding=padding, bias=False, mode='CRC', negative_slope=0.2))
for _ in range(15):
layers.append(
nn.Conv1d(in_channels=features, out_channels=features, kernel_size=kernel_size, padding=padding,
bias=False))
layers.append(nn.BatchNorm1d(features))
layers.append(nn.ReLU(inplace=True))
# layers.append(B.ResBlock(features, features, kernel_size=kernel_size, stride=1, padding=padding, bias=False, mode='CRC', negative_slope=0.2))
layers.append(nn.Conv1d(in_channels=features, out_channels=channels, kernel_size=kernel_size, padding=padding,
bias=False)) # 原来out_channels为channels
L = []
L.append(nn.Conv1d(in_channels=channels, out_channels=features, kernel_size=kernel_size, padding=padding,
bias=False))
L.append(nn.BatchNorm1d(features))
L.append(nn.ReLU(inplace=True))
for i in range(7):
L.append(nn.Conv1d(in_channels=features, out_channels=features, kernel_size=kernel_size, padding=2, groups=groups,
bias=False, dilation=2))
L.append(nn.ReLU(inplace=True))
L.append(nn.Conv1d(in_channels=features, out_channels=features, kernel_size=kernel_size, padding=padding,
bias=False))
L.append(nn.BatchNorm1d(features))
L.append(nn.ReLU(inplace=True))
for i in range(6):
L.append(nn.Conv1d(in_channels=features, out_channels=features, kernel_size=kernel_size, padding=2,
groups=groups,bias=False, dilation=2))
L.append(nn.ReLU(inplace=True))
L.append(nn.Conv1d(in_channels=features, out_channels=features, kernel_size=kernel_size, padding=padding,
bias=False))
L.append(nn.BatchNorm1d(features))
L.append(nn.ReLU(inplace=True))
L.append(nn.Conv1d(in_channels=features, out_channels=channels, kernel_size=kernel_size, padding=padding,
bias=False))
self.BRDNet_first = nn.Sequential(*layers)
self.BRDNet_second = nn.Sequential(*L)
self.conv1 = nn.Sequential(
nn.Conv1d(in_channels=channels*2, out_channels=channels, kernel_size=kernel_size, padding=padding,
groups=groups, bias=False))
def forward(self, x):
out1 = self.BRDNet_first(x)
out2 = self.BRDNet_second(x)
out1 = x - out1
out2 = x - out2
out = torch.cat([out1, out2], 1)
out = self.conv1(out)
out = x - out
return out
微信赞赏
支付宝赞赏