【时间】2020.01.07
【题目】《Perceptual Losses for Real-Time Style Transfer and Super-Resolution》论文及代码实现
- 《基于感知损失函数的实时风格转换和超分辨率重建》
- 文章地址:《Perceptual Losses for Real-Time Style Transfer and Super-Resolution》 arXiv.1603.08155
- 论文中文翻译:[译]
- 论文代码实现:GitHub代码 dxyang/StyleTransfer
一、论文要点
1、网络结构
- 包含图片转移网络和Loss网络两部分,包含style loss和 content loss 两种损失,基本思路和Neural Style transfer相似,只是使用了图片转移网络进行风格迁移,而Neural Style transfer是通过梯度上升直接修改content 图像x。
- Loss网络使用VGG-16,用于计算两个loss
- 图片转移网络的结构如下(包括3层卷积,5个残差block,2个反卷积(上采样))--参考上面的GitHub代码:
#pytorch
class ImageTransformNet(nn.Module):
def __init__(self):
super(ImageTransformNet, self).__init__()
# nonlineraity
self.relu = nn.ReLU()
self.tanh = nn.Tanh()
# encoding layers
self.conv1 = ConvLayer(3, 32, kernel_size=9, stride=1)
self.in1_e = nn.InstanceNorm2d(32, affine=True)
self.conv2 = ConvLayer(32, 64, kernel_size=3, stride=2)
self.in2_e = nn.InstanceNorm2d(64, affine=True)
self.conv3 = ConvLayer(64, 128, kernel_size=3, stride=2)
self.in3_e = nn.InstanceNorm2d(128, affine=True)
# residual layers
self.res1 = ResidualBlock(128)
self.res2 = ResidualBlock(128)
self.res3 = ResidualBlock(128)
self.res4 = ResidualBlock(128)
self.res5 = ResidualBlock(128)
# decoding layers
self.deconv3 = UpsampleConvLayer(128, 64, kernel_size=3, stride=1, upsample=2 )
self.in3_d = nn.InstanceNorm2d(64, affine=True)
self.deconv2 = UpsampleConvLayer(64, 32, kernel_size=3, stride=1, upsample=2 )
self.in2_d = nn.InstanceNorm2d(32, affine=True)
self.deconv1 = UpsampleConvLayer(32, 3, kernel_size=9, stride=1)
self.in1_d = nn.InstanceNorm2d(3, affine=True)
def forward(self, x):
# encode
y = self.relu(self.in1_e(self.conv1(x)))
y = self.relu(self.in2_e(self.conv2(y)))
y = self.relu(self.in3_e(self.conv3(y)))
# residual layers
y = self.res1(y)
y = self.res2(y)
y = self.res3(y)
y = self.res4(y)
y = self.res5(y)
# decode
y = self.relu(self.in3_d(self.deconv3(y)))
y = self.relu(self.in2_d(self.deconv2(y)))
#y = self.tanh(self.in1_d(self.deconv1(y)))
y = self.deconv1(y)
return y
- 训练时使用1张风格图片和多张内容图片进行训练,更新图片转移网络,训练完成后图片转移网络能够直接对输入的内容图片进行风格迁移。