Non-local(非局部平均)
1.背景
图像去噪是非常基础也是非常必要的研究,去噪常常在更高级的图像处理之前进行,是图像处理的基础。图像中的噪声常常用高斯噪声N(μ,σ^2)来近似表示。 一个有效的去除高斯噪声的方式是图像求平均,对N幅相同的图像求平均的结果将使得高斯噪声的方差降低到原来的N分之一,现在效果比较好的去噪算法都是基于这一思想来进行算法设计。(此处附上冈萨雷斯数字图像处理一段话)
通常的CNN网络模拟人的认知过程,在网络的相邻两层之间使用局部连接来获取图像的局部特性,一般认为人对外界的认知是从局部到全局的,而图像的空间联系也是局部的像素联系较为紧密,而距离较远的像素相关性则较弱。因而,每个神经元其实没有必要对全局图像进行感知,只需要对局部进行感知,然后在更高层将局部的信息综合起来就得到了全局的信息。网络部分连通的思想,也是受启发于生物学里面的视觉系统结构,底层的去捕捉轮廓信息,中层的组合轮廓信息,高层的组合全局信息,最终不同的全局信息最终被综合,但由于采样以及信息逐层传递损失了大量信息,所以传统cnn在全局信息捕捉上存在局限性。
2.非局部神经网络(Non-local Neural Networks)
非局域化方法在计算某一个位置的响应的时候,是通过对所有位置的特征进行加权求和得到的。
上面的公式中,输入是x,输出是y,i和j分别代表输入的某个空间位置,x_i是一个向量,维数跟x的channel数一样,f是一个计算任意两点相似关系的函数,g是一个映射函数,将一个点映射成一个向量,可以看成是计算一个点的特征。也就是说,为了计算输出层的一个点,需要将输入的每个点都考虑一遍,而且考虑的方式很像attention:输出的某个点在原图上的attention,而mask则是相似性给出。参看下图。
相似度测量函数f:
为了能让non-local操作作为一个组件,可以直接插入任意的神经网络中,作者把non-local设计成residual block的形式,让non-local操作去学x的residual:
3.代码(以手写数字识别为例)
3.1 non_local_embedded_gaussia
class _NonLocalBlockND(nn.Module):
def __init__(self, in_channels, inter_channels=None, dimension=3, sub_sample=True, bn_layer=True):
super(_NonLocalBlockND, self).__init__()
assert dimension in [1, 2, 3]
self.dimension = dimension
self.sub_sample = sub_sample
self.in_channels = in_channels
self.inter_channels = inter_channels
if self.inter_channels is None:
self.inter_channels = in_channels // 2
if self.inter_channels == 0:
self.inter_channels = 1
if dimension == 3:
conv_nd = nn.Conv3d
max_pool_layer = nn.MaxPool3d(kernel_size=(1, 2, 2))
bn = nn.BatchNorm3d
elif dimension == 2:
conv_nd = nn.Conv2d
max_pool_layer = nn.MaxPool2d(kernel_size=(2, 2))
bn = nn.BatchNorm2d
else:
conv_nd = nn.Conv1d
max_pool_layer = nn.MaxPool1d(kernel_size=(2))
bn = nn.BatchNorm1d
self.g = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels,
kernel_size=1, stride=1, padding=0)
if bn_layer:
self.W = nn.Sequential(
conv_nd(in_channels=self.inter_channels, out_channels=self.in_channels,
kernel_size=1, stride=1, padding=0),
bn(self.in_channels)
)
nn.init.constant_(self.W[1].weight, 0)
nn.init.constant_(self.W[1].bias, 0)
else:
self.W = conv_nd(in_channels=self.inter_channels, out_channels=self.in_channels,
kernel_size=1, stride=1, padding=0)
nn.init.constant_(self.W.weight, 0)
nn.init.constant_(self.W.bias, 0)
self.theta = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels,
kernel_size=1, stride=1, padding=0)
self.phi = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels,
kernel_size=1, stride=1, padding=0)
if sub_sample:
self.g = nn.Sequential(self.g, max_pool_layer)
self.phi = nn.Sequential(self.phi, max_pool_layer)
def forward(self, x):
'''
:param x: (b, c, t, h, w)
:return:
'''
batch_size = x.size(0)
g_x = self.g(x).view(batch_size, self.inter_channels, -1)
g_x = g_x.permute(0, 2, 1)
theta_x = self.theta(x).view(batch_size, self.inter_channels, -1)
theta_x = theta_x.permute(0, 2, 1)
phi_x = self.phi(x).view(batch_size, self.inter_channels, -1)
f = torch.matmul(theta_x, phi_x)
f_div_C = F.softmax(f, dim=-1)
y = torch.matmul(f_div_C, g_x)
y = y.permute(0, 2, 1).contiguous()
y = y.view(batch_size, self.inter_channels, *x.size()[2:])
W_y = self.W(y)
z = W_y + x
return z
class NONLocalBlock1D(_NonLocalBlockND):
def __init__(self, in_channels, inter_channels=None, sub_sample=True, bn_layer=True):
super(NONLocalBlock1D, self).__init__(in_channels,
inter_channels=inter_channels,
dimension=1, sub_sample=sub_sample,
bn_layer=bn_layer)
class NONLocalBlock2D(_NonLocalBlockND):
def __init__(self, in_channels, inter_channels=None, sub_sample=True, bn_layer=True):
super(NONLocalBlock2D, self).__init__(in_channels,
inter_channels=inter_channels,
dimension=2, sub_sample=sub_sample,
bn_layer=bn_layer)
class NONLocalBlock3D(_NonLocalBlockND):
def __init__(self, in_channels, inter_channels=None, sub_sample=True, bn_layer=True):
super(NONLocalBlock3D, self).__init__(in_channels,
inter_channels=inter_channels,
dimension=3, sub_sample=sub_sample,
bn_layer=bn_layer)
if __name__ == '__main__':
import torch
for (sub_sample, bn_layer) in [(True, True), (False, False), (True, False), (False, True)]:
img = torch.zeros(2, 3, 20)
net = NONLocalBlock1D(3, sub_sample=sub_sample, bn_layer=bn_layer)
out = net(img)
print(out.size())
img = torch.zeros(2, 3, 20, 20)
net = NONLocalBlock2D(3, sub_sample=sub_sample, bn_layer=bn_layer)
out = net(img)
print(out.size())
img = torch.randn(2, 3, 8, 20, 20)
net = NONLocalBlock3D(3, sub_sample=sub_sample, bn_layer=bn_layer)
out = net(img)
print(out.size())
3.2 net
class Network(nn.Module):
def __init__(self):
super(Network, self).__init__()
self.convs = nn.Sequential(
nn.Conv2d(in_channels=1, out_channels=32, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(32),
nn.ReLU(),
nn.MaxPool2d(2),
NONLocalBlock2D(in_channels=32),
nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.MaxPool2d(2),
NONLocalBlock2D(in_channels=64),
nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(128),
nn.ReLU(),
nn.MaxPool2d(2),
)
self.fc = nn.Sequential(
nn.Linear(in_features=128*3*3, out_features=256),
nn.ReLU(),
nn.Dropout(0.5),
nn.Linear(in_features=256, out_features=10)
)
def forward(self, x):
batch_size = x.size(0)
output = self.convs(x).view(batch_size, -1)
output = self.fc(output)
return output
if __name__ == '__main__':
import torch
img = torch.randn(3, 1, 28, 28)
net = Network()
out = net(img)
print(out.size())
3.3 train
train_data = torchvision.datasets.MNIST(root='./mnist', train=True,
transform=torchvision.transforms.ToTensor(),
download=True)
test_data = torchvision.datasets.MNIST(root='./mnist/',
transform=torchvision.transforms.ToTensor(),
train=False)
train_loader = Data.DataLoader(dataset=train_data, batch_size=128, shuffle=True)
test_loader = Data.DataLoader(dataset=test_data, batch_size=128, shuffle=False)
train_batch_num = len(train_loader)
test_batch_num = len(test_loader)
net = Network()
if torch.cuda.is_available():
net = nn.DataParallel(net)
net.cuda()
opt = torch.optim.Adam(net.parameters(), lr=0.001)
loss_func = nn.CrossEntropyLoss()
for epoch_index in range(10):
st = time.time()
torch.set_grad_enabled(True)
net.train()
for train_batch_index, (img_batch, label_batch) in enumerate(train_loader):
if torch.cuda.is_available():
img_batch = img_batch.cuda()
label_batch = label_batch.cuda()
predict = net(img_batch)
loss = loss_func(predict, label_batch)
net.zero_grad()
loss.backward()
opt.step()
print('(LR:%f) Time of a epoch:%.4fs' % (opt.param_groups[0]['lr'], time.time()-st))
torch.set_grad_enabled(False)
net.eval()
total_loss = []
total_acc = 0
total_sample = 0
for test_batch_index, (img_batch, label_batch) in enumerate(test_loader):
if torch.cuda.is_available():
img_batch = img_batch.cuda()
label_batch = label_batch.cuda()
predict = net(img_batch)
loss = loss_func(predict, label_batch)
predict = predict.argmax(dim=1)
acc = (predict == label_batch).sum()
total_loss.append(loss)
total_acc += acc
total_sample += img_batch.size(0)
net.train()
mean_acc = total_acc.item() * 1.0 / total_sample
mean_loss = sum(total_loss) / total_loss.__len__()
print('[Test] epoch[%d/%d] acc:%.4f%% loss:%.4f\n'
% (epoch_index, 10, mean_acc * 100, mean_loss.item()))