图像分割之U-Net

利用U-Net处理图像分割问题

图像分割应用场景

常用于医学上的影像分割

数据

标签

U-Net

网络优化

网络采用了最大池化进行下采样，一般的我们认为maxpool、padding会破坏图片的位置信息，所以可以用卷积层代替池化层，这样有利于保护位置信息，并且增加网络容量，而上采样不用转置卷积，因为转置卷积会产生象棋格效应，这对于图像语义信息产生了影响，所以可以采用邻近插值法代替转置卷积，由于随着网络层次的加深会产生梯度弥散问题，所以加了skip connection，可以有效地缓解这一问题，并且skip connection可以让信息流动更加通畅，这也接下来U-Net的改进版本U-Net++的提出

skip connection

一般的像这种稠密信息转稠密信息，我们都可以把网络设计成类似于哑铃的结构(编解码结构)，图片经过下采样过后，留下的是图片的轮廓信息，比较抽象和高级，那么可以看成是总结，而加了skip connection可以看成是增加细节部分，这也是U-Net比传统卷据FCN好的地方，而到了U-Net++作者大量的使用了skip connection这种残差结构

U-Net++

利用pytorch实现U-Net

import torch.nn as nn
import torch
from torch.nn import functional

#把常用的2个卷积操作简单封装下
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,stride=1),
            nn.BatchNorm2d(out_ch), #添加了BN层
            nn.ReLU(inplace=True),
            nn.Conv2d(out_ch, out_ch, 3, padding=1,stride=1),
            nn.BatchNorm2d(out_ch),
            nn.ReLU(inplace=True)
        )

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

class DownSample(nn.Module):
    def __init__(self,in_ch,out_ch):
        super(DownSample, self).__init__()
        self.downsample = nn.Sequential(
            nn.Conv2d(in_ch,out_ch,2,padding=0,stride=2)
        )
    def forward(self,x):
        return self.downsample(x)

class UpSample(nn.Module):
    def __init__(self):
        super(UpSample, self).__init__()

    def forward(self,x):
        return functional.interpolate(x,scale_factor=2,mode='nearest')

class Unet(nn.Module):
    def __init__(self, in_ch=1, out_ch=1):
        super(Unet, self).__init__()
        self.conv1 = DoubleConv(in_ch, 64)
        self.pool1 = DownSample(64,64)
        self.conv2 = DoubleConv(64, 128)
        self.pool2 = DownSample(128,128)
        self.conv3 = DoubleConv(128, 256)
        self.pool3 = DownSample(256,256)
        self.conv4 = DoubleConv(256, 512)
        self.pool4 = DownSample(512,512)
        self.conv5 = DoubleConv(512, 1024)
        # 逆卷积，也可以使用上采样
        self.up6 = UpSample()#nn.ConvTranspose2d(1024, 512, 2, stride=2)
        self.conv6 = DoubleConv(1024+512, 512)
        self.up7 = UpSample()
        self.conv7 = DoubleConv(512+256, 256)
        self.up8 = UpSample()
        self.conv8 = DoubleConv(256+128, 128)
        self.up9 = UpSample()
        self.conv9 = DoubleConv(128+64, 64)
        self.conv10 = nn.Conv2d(64, out_ch, 1)

    def forward(self, x):
        c1 = self.conv1(x)
        # print(c1.shape)
        p1 = self.pool1(c1)
        # print(p1.shape)
        c2 = self.conv2(p1)
        # print(c2.shape)
        p2 = self.pool2(c2)
        # print(p2.shape)
        c3 = self.conv3(p2)
        # print(c3.shape)
        p3 = self.pool3(c3)
        # print(p3.shape)
        c4 = self.conv4(p3)
        # print(c4.shape)
        p4 = self.pool4(c4)
        # print(p4.shape)
        c5 = self.conv5(p4)
        # print(c5.shape)
        up_6 = self.up6(c5)
        # print(up_6.shape)
        # print(c4.shape)
        merge6 = torch.cat([up_6, c4], dim=1)
        # print(merge6.shape)
        c6 = self.conv6(merge6)
        # print(c6.shape)
        up_7 = self.up7(c6)
        # print(up_7.shape)
        merge7 = torch.cat([up_7, c3], dim=1)
        c7 = self.conv7(merge7)
        # print(c7.shape)
        up_8 = self.up8(c7)
        # print(up_8.shape)
        merge8 = torch.cat([up_8, c2], dim=1)
        c8 = self.conv8(merge8)
        # print(c8.shape)
        up_9 = self.up9(c8)
        # print(up_9.shape)
        merge9 = torch.cat([up_9, c1], dim=1)
        c9 = self.conv9(merge9)
        # print(c9.shape)
        c10 = self.conv10(c9)
        # print(c10.shape)
        out = nn.Sigmoid()(c10)
        return out
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104

效果图