import torch import torch.nn as nn import torch.nn.functional as F import functools # import arch_util as arch_util # from NAFBlock import * import kornia import torch.nn.functional as F import torchvision.models try: import archs.arch_util as arch_util from archs.NAFBlock import * except: import arch_util as arch_util from NAFBlock import * class VGG19(torch.nn.Module): def __init__(self, requires_grad=False): super().__init__() vgg_pretrained_features = torchvision.models.vgg19(pretrained=True).features self.slice1 = torch.nn.Sequential() self.slice2 = torch.nn.Sequential() self.slice3 = torch.nn.Sequential() self.slice4 = torch.nn.Sequential() self.slice5 = torch.nn.Sequential() for x in range(2): self.slice1.add_module(str(x), vgg_pretrained_features[x]) for x in range(2, 7): self.slice2.add_module(str(x), vgg_pretrained_features[x]) for x in range(7, 12): self.slice3.add_module(str(x), vgg_pretrained_features[x]) for x in range(12, 21): self.slice4.add_module(str(x), vgg_pretrained_features[x]) for x in range(21, 30): self.slice5.add_module(str(x), vgg_pretrained_features[x]) if not requires_grad: for param in self.parameters(): param.requires_grad = False def forward(self, X): h_relu1 = self.slice1(X) h_relu2 = self.slice2(h_relu1) h_relu3 = self.slice3(h_relu2) h_relu4 = self.slice4(h_relu3) h_relu5 = self.slice5(h_relu4) out = [h_relu1, h_relu2, h_relu3, h_relu4, h_relu5] return out class VGGLoss(nn.Module): def __init__(self): super(VGGLoss, self).__init__() self.vgg = VGG19().cuda() # self.criterion = nn.L1Loss() self.criterion = nn.L1Loss(reduction='sum') self.criterion2 = nn.L1Loss() self.weights = [1.0 / 32, 1.0 / 16, 1.0 / 8, 1.0 / 4, 1.0] def forward(self, x, y): x_vgg, y_vgg = self.vgg(x), self.vgg(y) # print(x_vgg.shape, x_vgg.dtype, torch.max(x_vgg), torch.min(x_vgg), y_vgg.shape, y_vgg.dtype, torch.max(y_vgg), torch.min(y_vgg)) loss = 0 for i in range(len(x_vgg)): # print(x_vgg[i].shape, y_vgg[i].shape, 'hey') loss += self.weights[i] * self.criterion(x_vgg[i], y_vgg[i].detach()) # print(loss, i, 'hey') return loss class FourNet(nn.Module): def __init__(self, nf=64): super(FourNet, self).__init__() # AMPLITUDE ENHANCEMENT self.AmpNet = nn.Sequential( AmplitudeNet_skip(8), nn.Sigmoid() ) self.nf = nf ResidualBlock_noBN_f = functools.partial(arch_util.ResidualBlock_noBN, nf=nf) self.conv_first_1 = nn.Conv2d(3 * 2, nf, 3, 1, 1, bias=True) self.conv_first_2 = nn.Conv2d(nf, nf, 3, 2, 1, bias=True) self.conv_first_3 = nn.Conv2d(nf, nf, 3, 2, 1, bias=True) self.feature_extraction = arch_util.make_layer(ResidualBlock_noBN_f, 1) self.recon_trunk = arch_util.make_layer(ResidualBlock_noBN_f, 1) self.upconv1 = nn.Conv2d(nf*2, nf * 4, 3, 1, 1, bias=True) self.upconv2 = nn.Conv2d(nf*2, nf * 4, 3, 1, 1, bias=True) self.pixel_shuffle = nn.PixelShuffle(2) self.HRconv = nn.Conv2d(nf*2, nf, 3, 1, 1, bias=True) self.conv_last = nn.Conv2d(nf, 3, 3, 1, 1, bias=True) self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True) self.transformer = SFNet(nf, n = 4) self.recon_trunk_light = arch_util.make_layer(ResidualBlock_noBN_f, 6) def get_mask(self,dark): # SNR map light = kornia.filters.gaussian_blur2d(dark, (5, 5), (1.5, 1.5)) dark = dark[:, 0:1, :, :] * 0.299 + dark[:, 1:2, :, :] * 0.587 + dark[:, 2:3, :, :] * 0.114 light = light[:, 0:1, :, :] * 0.299 + light[:, 1:2, :, :] * 0.587 + light[:, 2:3, :, :] * 0.114 noise = torch.abs(dark - light) mask = torch.div(light, noise + 0.0001) batch_size = mask.shape[0] height = mask.shape[2] width = mask.shape[3] mask_max = torch.max(mask.view(batch_size, -1), dim=1)[0] mask_max = mask_max.view(batch_size, 1, 1, 1) mask_max = mask_max.repeat(1, 1, height, width) mask = mask * 1.0 / (mask_max + 0.0001) mask = torch.clamp(mask, min=0, max=1.0) return mask.float() def forward(self, x): # AMPLITUDE ENHANCEMENT #--------------------------------------------------------Frequency Stage--------------------------------------------------- _, _, H, W = x.shape image_fft = torch.fft.fft2(x, norm='backward') mag_image = torch.abs(image_fft) pha_image = torch.angle(image_fft) curve_amps = self.AmpNet(x) mag_image = mag_image / (curve_amps + 0.00000001) # * d4 real_image_enhanced = mag_image * torch.cos(pha_image) imag_image_enhanced = mag_image * torch.sin(pha_image) img_amp_enhanced = torch.fft.ifft2(torch.complex(real_image_enhanced, imag_image_enhanced), s=(H, W), norm='backward').real x_center = img_amp_enhanced rate = 2 ** 3 pad_h = (rate - H % rate) % rate pad_w = (rate - W % rate) % rate if pad_h != 0 or pad_w != 0: x_center = F.pad(x_center, (0, pad_w, 0, pad_h), "reflect") x = F.pad(x, (0, pad_w, 0, pad_h), "reflect") #------------------------------------------Spatial Stage--------------------------------------------------------------------- L1_fea_1 = self.lrelu(self.conv_first_1(torch.cat((x_center,x),dim=1))) L1_fea_2 = self.lrelu(self.conv_first_2(L1_fea_1)) # Encoder L1_fea_3 = self.lrelu(self.conv_first_3(L1_fea_2)) fea = self.feature_extraction(L1_fea_3) fea_light = self.recon_trunk_light(fea) h_feature = fea.shape[2] w_feature = fea.shape[3] mask_image = self.get_mask(x_center) # SNR Map mask = F.interpolate(mask_image, size=[h_feature, w_feature], mode='nearest') # Resize and Normalize SNR map fea_unfold = self.transformer(fea) channel = fea.shape[1] mask = mask.repeat(1, channel, 1, 1) fea = fea_unfold * (1 - mask) + fea_light * mask # SNR-based Interaction out_noise = self.recon_trunk(fea) out_noise = torch.cat([out_noise, L1_fea_3], dim=1) out_noise = self.lrelu(self.pixel_shuffle(self.upconv1(out_noise))) out_noise = torch.cat([out_noise, L1_fea_2], dim=1) # Decoder out_noise = self.lrelu(self.pixel_shuffle(self.upconv2(out_noise))) out_noise = torch.cat([out_noise, L1_fea_1], dim=1) out_noise = self.lrelu(self.HRconv(out_noise)) out_noise = self.conv_last(out_noise) out_noise = out_noise + x out_noise = out_noise[:, :, :H, :W] return out_noise, mag_image, x_center, mask_image