基于UNet的变化检测

基于UNet的变化检测

论文名称：基于改进 UNet 孪生网络的遥感影像矿区变化检测
论文地址：http://www.chinacaj.net/i,2,425089,0.html

一、内容框架：

1、改进的UNet孪生网络结构

2、影响中心环绕

3、特征金字塔

二、网络框架复现 (Pytorch)

import torch.nn as nn
import torch.utils.model_zoo as model_zoo
from torch.nn import functional as F
import torch

        
class DoubleConv(nn.Module):
    def __init__(self, in_ch, out_ch):
        super(DoubleConv, self).__init__()
        self.conv = nn.Sequential(
            nn.Conv2d(in_ch, out_ch, 3, padding=1),
            nn.BatchNorm2d(out_ch), #添加了BN层
            nn.ReLU(inplace=True),
            nn.Conv2d(out_ch, out_ch, 3, padding=1),
            nn.BatchNorm2d(out_ch),
            nn.ReLU(inplace=True)
        )

    def forward(self, input):
        return self.conv(input)

class Unet(nn.Module):
    def __init__(self, in_ch, out_ch):
        super(Unet, self).__init__()
        self.pool = nn.MaxPool2d(2)
        self.conv1 = DoubleConv(in_ch, 64)
        self.pool1 = nn.MaxPool2d(2)
        self.conv2 = DoubleConv(64, 128)
        self.pool2 = nn.MaxPool2d(2)
        self.conv3 = DoubleConv(128, 256)
        self.pool3 = nn.MaxPool2d(2)
        self.conv4 = DoubleConv(256, 512)
        self.pool4 = nn.MaxPool2d(2)
        self.conv5 = DoubleConv(512, 1024)
        # 逆卷积，也可以使用上采样(保证k=stride,stride即上采样倍数)
        self.up6 = nn.ConvTranspose2d(1024, 512, 2, stride=2)
        self.conv6 = DoubleConv(1536, 512)
        self.up7 = nn.ConvTranspose2d(512, 256, 2, stride=2)
        self.conv7 = DoubleConv(768, 256)
        self.up8 = nn.ConvTranspose2d(256, 128, 2, stride=2)
        self.conv8 = DoubleConv(384, 128)
        self.up9 = nn.ConvTranspose2d(128, 64, 2, stride=2)
        self.conv9 = DoubleConv(192, 64)
        self.conv10 = nn.Conv2d(64, out_ch, 1)
        self.conv1_dilation = nn.Conv2d(2048, 256, 1, stride=1, padding=0, bias=False, dilation=1)  # dilation就是空洞率，即间隔
        self.conv2_dilation = nn.Conv2d(2048, 256, 2, stride=1, padding=2, bias=False, dilation=2)  # dilation就是空洞率，即间隔
        self.conv4_dilation = nn.Conv2d(2048, 256, 4, stride=1, padding=4, bias=False, dilation=4)  # dilation就是空洞率，即间隔
        self.global_pool = nn.AdaptiveAvgPool2d((1, 1)) 
        self.upsample = nn.Upsample(scale_factor=7, mode='bicubic', align_corners=True) 
        self.conv_c = nn.Conv2d(2816, 1024, 1, stride=1, padding=0, bias=False, dilation=1)  # dilation就是空洞率，即间隔
        self.upsample1 = nn.Upsample(scale_factor=2, mode='bicubic', align_corners=True) 

    def forward(self, A2016, A2019):
        A2016_resize = self.pool(A2016)
        A2019_resize = self.pool(A2019)
        A2016_crop = A2016[:,:,56:168,56:168]
        A2019_crop = A2019[:,:,56:168,56:168]
        c1_A2016_resize = self.conv1(A2016_resize)  # [2, 64, 112, 112] 
        c1_A2019_resize = self.conv1(A2019_resize)  # [2, 64, 112, 112]
        c1_A2016_crop = self.conv1(A2016_crop)  # [2, 64, 112, 112] 
        c1_A2019_crop = self.conv1(A2019_crop)  # [2, 64, 112, 112]
        c1_resize = torch.abs(torch.sub(c1_A2016_resize,c1_A2019_resize))
        c1_crop = torch.abs(torch.sub(c1_A2016_crop,c1_A2019_crop))
        c1 = torch.cat([c1_resize,c1_crop], dim = 1)

        p1_A2016_resize = self.pool1(c1_A2016_resize) # [2, 64, 56, 56]
        p1_A2019_resize = self.pool1(c1_A2019_resize) # [2, 64, 56, 56]
        p1_A2016_crop = self.pool1(c1_A2016_crop) # [2, 64, 56, 56]
        p1_A2019_crop = self.pool1(c1_A2019_crop) # [2, 64, 56, 56]
        c2_A2016_resize = self.conv2(p1_A2016_resize) # [2, 128, 56, 56]
        c2_A2019_resize = self.conv2(p1_A2019_resize) # [2, 128, 56, 56]
        c2_A2016_crop = self.conv2(p1_A2016_crop) # [2, 128, 56, 56]
        c2_A2019_crop = self.conv2(p1_A2019_crop) # [2, 128, 56, 56]
        c2_resize = torch.abs(torch.sub(c2_A2016_resize,c2_A2019_resize))
        c2_crop = torch.abs(torch.sub(c2_A2016_crop,c2_A2019_crop))
        c2 = torch.cat([c2_resize,c2_crop], dim = 1)

        p2_A2016_resize = self.pool2(c2_A2016_resize) # [2, 128, 28, 28]
        p2_A2019_resize = self.pool2(c2_A2019_resize) # [2, 128, 28, 28]
        p2_A2016_crop = self.pool2(c2_A2016_crop) # [2, 128, 28, 28]
        p2_A2019_crop = self.pool2(c2_A2019_crop) # [2, 128, 28, 28]
        c3_A2016_resize = self.conv3(p2_A2016_resize) # [2, 256, 28, 28]
        c3_A2019_resize = self.conv3(p2_A2019_resize) # [2, 256, 28, 28]
        c3_A2016_crop = self.conv3(p2_A2016_crop) # [2, 256, 28, 28]
        c3_A2019_crop = self.conv3(p2_A2019_crop) # [2, 256, 28, 28]
        c3_resize = torch.abs(torch.sub(c3_A2016_resize,c3_A2019_resize))
        c3_crop = torch.abs(torch.sub(c3_A2016_crop,c3_A2019_crop))
        c3 = torch.cat([c3_resize,c3_crop], dim = 1)


        p3_A2016_resize = self.pool3(c3_A2016_resize) # [2, 256, 14, 14]
        p3_A2019_resize = self.pool3(c3_A2019_resize) # [2, 256, 14, 14]
        p3_A2016_crop = self.pool3(c3_A2016_crop) # [2, 256, 14, 14]
        p3_A2019_crop = self.pool3(c3_A2019_crop) # [2, 256, 14, 14]
        c4_A2016_resize = self.conv4(p3_A2016_resize) # [2, 512, 14, 14]
        c4_A2019_resize = self.conv4(p3_A2019_resize) # [2, 512, 14, 14]
        c4_A2016_crop = self.conv4(p3_A2016_crop) # [2, 512, 14, 14]
        c4_A2019_crop = self.conv4(p3_A2019_crop) # [2, 512, 14, 14]
        c4_resize = torch.abs(torch.sub(c4_A2016_resize,c4_A2019_resize))
        c4_crop = torch.abs(torch.sub(c4_A2016_crop,c4_A2019_crop))
        c4 = torch.cat([c4_resize,c4_crop], dim = 1)

        p4_A2016_resize = self.pool4(c4_A2016_resize) # [2, 512, 7, 7]
        p4_A2019_resize = self.pool4(c4_A2019_resize) # [2, 512, 7, 7]
        p4_A2016_crop = self.pool4(c4_A2016_crop) # [2, 512, 7, 7]
        p4_A2019_crop = self.pool4(c4_A2019_crop) # [2, 512, 7, 7]
        c5_A2016_resize = self.conv5(p4_A2016_resize) # [2, 1024, 7, 7]
        c5_A2019_resize = self.conv5(p4_A2019_resize) # [2, 1024, 7, 7]
        c5_A2016_crop = self.conv5(p4_A2016_crop) # [2, 1024, 7, 7]
        c5_A2019_crop = self.conv5(p4_A2019_crop) # [2, 1024, 7, 7]
        c5_resize = torch.abs(torch.sub(c5_A2016_resize,c5_A2019_resize))
        c5_crop = torch.abs(torch.sub(c5_A2016_crop,c5_A2019_crop))
        c5 = torch.cat([c5_resize,c5_crop], dim = 1)
        c5_1_dilation = self.conv1_dilation(c5)
        c5_2_dilation = self.conv1_dilation(c5)
        c5_4_dilation = self.conv1_dilation(c5)
        c5_AVG = self.global_pool(c5)
        c5 = self.upsample(c5_AVG)
        c5 = torch.cat([c5_1_dilation,c5_2_dilation,c5_4_dilation,c5], dim = 1)
        c5 = self.conv_c(c5)
         
        up_6 = self.up6(c5) # [2, 512, 14, 14]
        merge6 = torch.cat([up_6, c4], dim=1) # [2, 1024, 14, 14]
        c6 = self.conv6(merge6) # [2, 512, 14, 14]
        up_7 = self.up7(c6) # [2, 256, 28, 28]

        merge7 = torch.cat([up_7, c3], dim=1) # [2, 512, 28, 28]
        c7 = self.conv7(merge7) # [2, 256, 28, 28]
        up_8 = self.up8(c7) # [2, 128, 56, 56]

        merge8 = torch.cat([up_8, c2], dim=1) # [2, 256, 56, 56]
        c8 = self.conv8(merge8) # [2, 128, 56, 56]
        up_9 = self.up9(c8) # [2, 64, 112, 112]

        merge9 = torch.cat([up_9, c1], dim=1) # [2, 128, 112, 112]
        c9 = self.conv9(merge9) # [2, 64, 112, 112]
        c10 = self.conv10(c9) # [2, 64, 112, 112]
        c10 = self.upsample1(c10)
        out = nn.Sigmoid()(c10) # [2, 64, 112, 112]
        return out

if __name__ == "__main__":
    A2016 = torch.randn(2, 3, 224, 224)
    A2019 = torch.randn(2, 3, 224, 224)
    UNet = Unet(3,3)
    out_result = UNet(A2016,A2019)
    print(out_result)
    print(out_result.shape)