第一步:PSD介绍
以往的研究主要集中在具有合成模糊图像的训练模型上,当模型用于真实世界的模糊图像时,会导致性能下降。
为了解决上述问题,提高去雾的泛化性能,作者提出了一种Principled Synthetic-to-real Dehazing (PSD)框架。
本文提出的PSD适用于将现有的去雾模型推广到实际领域,包括两个阶段:有监督的预训练和无监督的微调。
预训练阶段,作者将选定的去雾模型主干修改为一个基于物理模型的网络,并用合成数据训练该网络。利用设计良好的主干,我们可以得到一个预先训练的模型,在合成域上具有良好的去雾性能。
微调阶段,作者利用真实的模糊图像以无监督的方式训练模型。
本文的贡献:
- 作者将真实世界的去雾任务重新定义为一个合成到真实的泛化框架:首先一个在合成配对数据上预先训练的去雾模型主干,真实的模糊图像随后将被利用以一种无监督的方式微调模型。PSD易于使用,可以以大多数深度去雾模型为骨干。
- 由于没有清晰的真实图像作为监督,作者利用几个流行的、有充分根据的物理先验来指导微调。作者将它们合并成一个预先的损失committee,作为具体任务的代理指导,这一部分也是PSD的核心。
- 性能达到SOTA
第二步:PSD网络结构
首先对两个框架大的方向做一个整体概述:
Pre-training
首先采用目前性能最好的框架之一作为网络的主干
然后我们将主干修改为一个基于物理的网络,根据一个单一的雾图同时生成干净的图像 J,传输图 t 和大气光 A,为了共同优化这三个分量,作者加入了一个重建损失,它引导网络输出服从物理散射模型。
在这个阶段,只使用标记的合成数据进行训练,最终得到一个在合成域上预训练的模型。
Fine-tuning
作者利用未标记的真实数据将预训练模型从合成域推广到真实域。受去雾强物理背景的启发,作者认为一个高质量的无雾图像应该遵循一些特定的统计规则,这些规则可以从图像先验中推导出来。此外,单一先验提供的物理知识并不总是可靠的,所以作者的目标是找到多个先验的组合,希望它们能够相互补充。
基于上述,作者设计了一个先验损失committee来作为任务特定的代理指导,用于训练未标记的真实数据。
此外,作者应用了一种learning without forgetting (LwF)的方法,该方法通过将原始任务的训练数据(即合成的模糊图像)通过网络运转到同真实的模糊数据一起,从而强行使得模型记忆合成领域的知识。
第三步:模型代码展示
import torch
import torch.nn as nn
import torch.nn.functional as Fclass BlockUNet1(nn.Module):def __init__(self, in_channels, out_channels, upsample=False, relu=False, drop=False, bn=True):super(BlockUNet1, self).__init__()self.conv = nn.Conv2d(in_channels, out_channels, 4, 2, 1, bias=False)self.deconv = nn.ConvTranspose2d(in_channels, out_channels, 4, 2, 1, bias=False)self.dropout = nn.Dropout2d(0.5)self.batch = nn.InstanceNorm2d(out_channels)self.upsample = upsampleself.relu = reluself.drop = dropself.bn = bndef forward(self, x):if self.relu == True:y = F.relu(x)elif self.relu == False:y = F.leaky_relu(x, 0.2)if self.upsample == True:y = self.deconv(y)if self.bn == True:y = self.batch(y)if self.drop == True:y = self.dropout(y)elif self.upsample == False:y = self.conv(y)if self.bn == True:y = self.batch(y)if self.drop == True:y = self.dropout(y)return yclass G2(nn.Module):def __init__(self, in_channels, out_channels):super(G2, self).__init__()self.conv = nn.Conv2d(in_channels, 8, 4, 2, 1, bias=False)self.layer1 = BlockUNet1(8, 16)self.layer2 = BlockUNet1(16, 32)self.layer3 = BlockUNet1(32, 64)self.layer4 = BlockUNet1(64, 64)self.layer5 = BlockUNet1(64, 64)self.layer6 = BlockUNet1(64, 64)self.layer7 = BlockUNet1(64, 64)self.dlayer7 = BlockUNet1(64, 64, True, True, True, False)self.dlayer6 = BlockUNet1(128, 64, True, True, True)self.dlayer5 = BlockUNet1(128, 64, True, True, True)self.dlayer4 = BlockUNet1(128, 64, True, True)self.dlayer3 = BlockUNet1(128, 32, True, True)self.dlayer2 = BlockUNet1(64, 16, True, True)self.dlayer1 = BlockUNet1(32, 8, True, True)self.relu = nn.ReLU()self.dconv = nn.ConvTranspose2d(16, out_channels, 4, 2, 1, bias=False)self.lrelu = nn.LeakyReLU(0.2)def forward(self, x):y1 = self.conv(x)y2 = self.layer1(y1)y3 = self.layer2(y2)y4 = self.layer3(y3)y5 = self.layer4(y4)y6 = self.layer5(y5)y7 = self.layer6(y6)y8 = self.layer7(y7)dy8 = self.dlayer7(y8)concat7 = torch.cat([dy8, y7], 1)dy7 = self.dlayer6(concat7)concat6 = torch.cat([dy7, y6], 1)dy6 = self.dlayer5(concat6)concat5 = torch.cat([dy6, y5], 1)dy5 = self.dlayer4(concat5)concat4 = torch.cat([dy5, y4], 1)dy4 = self.dlayer3(concat4)concat3 = torch.cat([dy4, y3], 1)dy3 = self.dlayer2(concat3)concat2 = torch.cat([dy3, y2], 1)dy2 = self.dlayer1(concat2)concat1 = torch.cat([dy2, y1], 1)out = self.relu(concat1)out = self.dconv(out)out = self.lrelu(out)return F.avg_pool2d(out, (out.shape[2], out.shape[3]))def default_conv(in_channels, out_channels, kernel_size, bias=True):return nn.Conv2d(in_channels, out_channels, kernel_size,padding=(kernel_size//2), bias=bias)class PALayer(nn.Module):def __init__(self, channel):super(PALayer, self).__init__()self.pa = nn.Sequential(nn.Conv2d(channel, channel // 8, 1, padding=0, bias=True),nn.ReLU(inplace=True),nn.Conv2d(channel // 8, 1, 1, padding=0, bias=True),nn.Sigmoid())def forward(self, x):y = self.pa(x)return x * yclass CALayer(nn.Module):def __init__(self, channel):super(CALayer, self).__init__()self.avg_pool = nn.AdaptiveAvgPool2d(1)self.ca = nn.Sequential(nn.Conv2d(channel, channel // 8, 1, padding=0, bias=True),nn.ReLU(inplace=True),nn.Conv2d(channel // 8, channel, 1, padding=0, bias=True),nn.Sigmoid())def forward(self, x):y = self.avg_pool(x)y = self.ca(y)return x * yclass Block(nn.Module):def __init__(self, conv, dim, kernel_size,):super(Block, self).__init__()self.conv1=conv(dim, dim, kernel_size, bias=True)self.act1=nn.ReLU(inplace=True)self.conv2=conv(dim,dim,kernel_size,bias=True)self.calayer=CALayer(dim)self.palayer=PALayer(dim)def forward(self, x):res=self.act1(self.conv1(x))res=res+x res=self.conv2(res)res=self.calayer(res)res=self.palayer(res)res += x return res
class Group(nn.Module):def __init__(self, conv, dim, kernel_size, blocks):super(Group, self).__init__()modules = [ Block(conv, dim, kernel_size) for _ in range(blocks)]modules.append(conv(dim, dim, kernel_size))self.gp = nn.Sequential(*modules)def forward(self, x):res = self.gp(x)res += xreturn resclass FFANet(nn.Module):def __init__(self,gps,blocks,conv=default_conv):super(FFANet, self).__init__()self.gps=gpsself.dim=64kernel_size=3pre_process = [conv(3, self.dim, kernel_size)]assert self.gps==3self.g1= Group(conv, self.dim, kernel_size,blocks=blocks)self.g2= Group(conv, self.dim, kernel_size,blocks=blocks)self.g3= Group(conv, self.dim, kernel_size,blocks=blocks)self.ca=nn.Sequential(*[nn.AdaptiveAvgPool2d(1),nn.Conv2d(self.dim*self.gps,self.dim//16,1,padding=0),nn.ReLU(inplace=True),nn.Conv2d(self.dim//16, self.dim*self.gps, 1, padding=0, bias=True),nn.Sigmoid()])self.palayer=PALayer(self.dim)self.conv_J_1 = nn.Conv2d(64, 64, 3, 1, 1, bias=False)self.conv_J_2 = nn.Conv2d(64, 3, 3, 1, 1, bias=False)self.conv_T_1 = nn.Conv2d(64, 16, 3, 1, 1, bias=False)self.conv_T_2 = nn.Conv2d(16, 1, 3, 1, 1, bias=False)post_precess = [conv(self.dim, self.dim, kernel_size),conv(self.dim, 3, kernel_size)]self.pre = nn.Sequential(*pre_process)self.post = nn.Sequential(*post_precess)self.ANet = G2(3, 3)def forward(self, x1, x2=0, Val=False):x = self.pre(x1)res1=self.g1(x)res2=self.g2(res1)res3=self.g3(res2)w=self.ca(torch.cat([res1,res2,res3],dim=1))w=w.view(-1,self.gps,self.dim)[:,:,:,None,None]out=w[:,0,::]*res1+w[:,1,::]*res2+w[:,2,::]*res3out=self.palayer(out)out_J = self.conv_J_1(out)out_J = self.conv_J_2(out_J)out_J = out_J + x1out_T = self.conv_T_1(out)out_T = self.conv_T_2(out_T)if Val == False:out_A = self.ANet(x1)else:out_A = self.ANet(x2)out_I = out_T * out_J + (1 - out_T) * out_A#x=self.post(out)return out, out_J, out_T, out_A, out_Iif __name__ == "__main__":net=FFA(gps=3,blocks=19)print(net)
第四步:运行
第五步:整个工程的内容
代码的下载路径(新窗口打开链接):基于深度学习神经网络的AI图像PSD去雾系统源码
有问题可以私信或者留言,有问必答