import torch import math import torch.nn as nn import torch.nn.functional as F from typing import List class Convolution(nn.Module): def __init__(self, in_ch, out_ch): super(Convolution, self).__init__() self.conv = nn.Sequential( nn.Conv2d(in_ch, out_ch, 3, 1, 1), nn.BatchNorm2d(out_ch), nn.ReLU(inplace=True), nn.Conv2d(out_ch, out_ch, 3, 1, 1), nn.BatchNorm2d(out_ch), nn.ReLU(inplace=True) ) def forward(self, input): return self.conv(input) class Curvature(torch.nn.Module): def __init__(self, ratio): super(Curvature, self).__init__() weights = torch.tensor([[[[-1/16, 5/16, -1/16], [5/16, -1, 5/16], [-1/16, 5/16, -1/16]]]]) self.weight = torch.nn.Parameter(weights).cuda() self.ratio = ratio def forward(self, x): B, C, H, W = x.size() x_origin = x x = x.reshape(B*C,1,H,W) out = F.conv2d(x, self.weight) out = torch.abs(out) p = torch.sum(out, dim=-1) p = torch.sum(p, dim=-1) p=p.reshape(B, C) _, index = torch.topk(p, int(self.ratio*C), dim=1) selected = [] for i in range(x_origin.shape[0]): selected.append(torch.index_select(x_origin[i], dim=0, index=index[i]).unsqueeze(0)) selected = torch.cat(selected, dim=0) return selected class Entropy_Hist(nn.Module): def __init__(self, ratio, win_w=3, win_h=3): super(Entropy_Hist, self).__init__() self.win_w = win_w self.win_h = win_h self.ratio = ratio def calcIJ_new(self, img_patch): total_p = img_patch.shape[-1] * img_patch.shape[-2] if total_p % 2 != 0: tem = torch.flatten(img_patch, start_dim=-2, end_dim=-1) center_p = tem[:, :, :, int(total_p / 2)] mean_p = (torch.sum(tem, dim=-1) - center_p) / (total_p - 1) if torch.is_tensor(img_patch): return center_p * 100 + mean_p else: return (center_p, mean_p) else: print("modify patch size") def histc_fork(ij): BINS = 256 B, C = ij.shape N = 16 BB = B // N min_elem = ij.min() max_elem = ij.max() ij = ij.view(N, BB, C) def f(x): with torch.no_grad(): res = [] for e in x: res.append(torch.histc(e, bins=BINS, min=min_elem, max=max_elem)) return res futures : List[torch.jit.Future[torch.Tensor]] = [] for i in range(N): futures.append(torch.jit.fork(f, ij[i])) results = [] for future in futures: results += torch.jit.wait(future) with torch.no_grad(): out = torch.stack(results) return out def forward(self, img): with torch.no_grad(): B, C, H, W = img.shape ext_x = int(self.win_w / 2) # 考虑滑动窗口大小，对原图进行扩边，扩展部分长度 ext_y = int(self.win_h / 2) new_width = ext_x + W + ext_x # 新的图像尺寸 new_height = ext_y + H + ext_y # 使用nn.Unfold依次获取每个滑动窗口的内容 nn_Unfold=nn.Unfold(kernel_size=(self.win_w,self.win_h),dilation=1,padding=ext_x,stride=1) # 能够获取到patch_img，shape=(B,C*K*K,L),L代表的是将每张图片由滑动窗口分割成多少块---->28*28的图像，3*3的滑动窗口，分成了28*28=784块 x = nn_Unfold(img) # (B,C*K*K,L) x= x.view(B,C,3,3,-1).permute(0,1,4,2,3) # (B,C*K*K,L) ---> (B,C,L,K,K) ij = self.calcIJ_new(x).reshape(B*C, -1) # 计算滑动窗口内中心的灰度值和窗口内除了中心像素的灰度均值,(B,C,L,K,K)---> (B,C,L) ---> (B*C,L) fij_packed = self.histc_fork(ij) p = fij_packed / (new_width * new_height) h_tem = -p * torch.log(torch.clamp(p, min=1e-40)) / math.log(2) a = torch.sum(h_tem, dim=1) # 对所有二维熵求和，得到这张图的二维熵 H = a.reshape(B,C) _, index = torch.topk(H, int(self.ratio*C), dim=1) # Nx3 selected = [] for i in range(img.shape[0]): selected.append(torch.index_select(img[i], dim=0, index=index[i]).unsqueeze(0)) selected = torch.cat(selected, dim=0) return selected class Network(nn.Module): def __init__(self, in_ch=3, mode='ori', ratio=None): super(Network, self).__init__() self.mode = mode if self.mode == 'ori': self.ratio = [0,0] if self.mode == 'curvature': self.ratio = ratio self.ife1 = Curvature(self.ratio[0]) self.ife2 = Curvature(self.ratio[1]) if self.mode == 'entropy': self.ratio = ratio self.ife1 = Entropy_Hist(self.ratio[0]) self.ife2 = Entropy_Hist(self.ratio[1]) # ---- U-Net ---- self.conv1 = Convolution(in_ch, 64) self.pool1 = nn.MaxPool2d(2) # feature map = shape(m/2,n/2,64) self.conv2 = Convolution(64, 128) self.pool2 = nn.MaxPool2d(2) # feature map = shapem/4,n/4,128) self.conv3 = Convolution(128, 256) self.pool3 = nn.MaxPool2d(2) # feature map = shape(m/8,n/8,256) self.conv4 = Convolution(256, 512) self.pool4 = nn.MaxPool2d(2) # feature map = shape(m/16,n/16,512) self.conv5 = Convolution(512, 1024) # feature map = shape(m/16,n/16,1024) self.up_conv1 = nn.ConvTranspose2d(in_channels=1024, out_channels=512, kernel_size=2, stride=2, padding=0, output_padding=0) self.conv6 = Convolution(1024, 512) # feature map = shape(m/8,n/8,512) self.up_conv2 = nn.ConvTranspose2d(512, 256, 2, 2, 0, 0) self.conv7 = Convolution(int(256*(2+self.ratio[1])), 256) # feature map = shape(m/4,n/4,256） self.up_conv3 = nn.ConvTranspose2d(256, 128, 2, 2, 0, 0) self.conv8 = Convolution(int(128*(2+self.ratio[0])), 128) # feature map = shape(m/2,n/2,128） self.up_conv4 = nn.ConvTranspose2d(128, 64, 2, 2, 0, 0) self.conv9 = Convolution(128, 64) # feature map = shape(m,n,64) self.out_conv1 = nn.Conv2d(64, 1, 1, 1, 0) def forward(self, x): c1 = self.conv1(x) p1 = self.pool1(c1) c2 = self.conv2(p1) p2 = self.pool2(c2) c3 = self.conv3(p2) p3 = self.pool3(c3) c4 = self.conv4(p3) p4 = self.pool4(c4) c5 = self.conv5(p4) if self.mode != 'ori': c2 = torch.cat([c2, self.ife1(c2)]) c3 = torch.cat([c3, self.ife2(c3)]) up1 = self.up_conv1(c5) merge1 = torch.cat([up1, c4], dim=1) c6 = self.conv6(merge1) up2 = self.up_conv2(c6) merge2 = torch.cat([up2, c3], dim=1) c7 = self.conv7(merge2) up3 = self.up_conv3(c7) merge3 = torch.cat([up3, c2], dim=1) c8 = self.conv8(merge3) up4 = self.up_conv4(c8) merge4 = torch.cat([up4, c1], dim=1) c9 = self.conv9(merge4) S_g_pred = self.out_conv1(c9) return S_g_pred