【小白学习PyTorch教程】十九、基于torch实现UNet 图像分割模型

@Author：Runsen

在图像领域，除了分类，CNN 今天还用于更高级的问题，如图像分割、对象检测等。图像分割是计算机视觉中的一个过程，其中图像被分割成代表图像中每个不同类别的不同段。

上面图片一段代表猫，另一段代表背景。

从自动驾驶汽车到卫星，图像分割在许多领域都很有用。其中最重要的是医学成像。

UNet 是一种卷积神经网络架构，在 CNN 架构几乎没有变化的情况下进行了扩展。它的发明是为了处理生物医学图像，其目标不仅是对是否存在感染进行分类，而且还要识别感染区域。

UNet

论文：https://arxiv.org/abs/1505.04597

UNet结构看起来像一个“U”，该架构由三部分组成：收缩部分、瓶颈部分和扩展部分。收缩段由许多收缩块组成。每个块接受一个输入，应用两个 3X3 卷积层，然后是 2X2 最大池化。每个块之后的内核或特征图的数量加倍，以便架构可以有效地学习复杂的结构。最底层介于收缩层和膨胀层之间。它使用两个 3X3 CNN 层，然后是 2X2 上卷积层。

每个块将输入传递给两个 3X3 CNN 层，然后是一个 2X2 上采样层。同样在每个块之后，卷积层使用的特征图数量减半以保持对称性。然而，每次输入也会附加相应收缩层的特征图。此操作将确保在收缩图像时学习的特征将用于重建它。扩展块的数量与收缩块的数量相同。之后，生成的映射通过另一个 3X3 CNN 层，特征映射的数量等于所需的片段数量。

torch实现

使用的数据集是：https://www.kaggle.com/paultimothymooney/chiu-2015

这个数据集用于分割糖尿病性黄斑水肿的光学相干断层扫描图像的图像。

对于mat的数据，使用scipy.io.loadmat进行加载

下面使用 Pytorch 框架实现了 UNet 模型，代码来源下面的Github：https://github.com/Hsankesara/DeepResearch

import torch
from torch import nn
import torch.nn.functional as F
import torch.optim as optimclass UNet(nn.Module):def contracting_block(self, in_channels, out_channels, kernel_size=3):block = torch.nn.Sequential(torch.nn.Conv2d(kernel_size=kernel_size, in_channels=in_channels, out_channels=out_channels),torch.nn.ReLU(),torch.nn.BatchNorm2d(out_channels),torch.nn.Conv2d(kernel_size=kernel_size, in_channels=out_channels, out_channels=out_channels),torch.nn.ReLU(),torch.nn.BatchNorm2d(out_channels),)return blockdef expansive_block(self, in_channels, mid_channel, out_channels, kernel_size=3):block = torch.nn.Sequential(torch.nn.Conv2d(kernel_size=kernel_size, in_channels=in_channels, out_channels=mid_channel),torch.nn.ReLU(),torch.nn.BatchNorm2d(mid_channel),torch.nn.Conv2d(kernel_size=kernel_size, in_channels=mid_channel, out_channels=mid_channel),torch.nn.ReLU(),torch.nn.BatchNorm2d(mid_channel),torch.nn.ConvTranspose2d(in_channels=mid_channel, out_channels=out_channels, kernel_size=3, stride=2, padding=1, output_padding=1))return  blockdef final_block(self, in_channels, mid_channel, out_channels, kernel_size=3):block = torch.nn.Sequential(torch.nn.Conv2d(kernel_size=kernel_size, in_channels=in_channels, out_channels=mid_channel),torch.nn.ReLU(),torch.nn.BatchNorm2d(mid_channel),torch.nn.Conv2d(kernel_size=kernel_size, in_channels=mid_channel, out_channels=mid_channel),torch.nn.ReLU(),torch.nn.BatchNorm2d(mid_channel),torch.nn.Conv2d(kernel_size=kernel_size, in_channels=mid_channel, out_channels=out_channels, padding=1),torch.nn.ReLU(),torch.nn.BatchNorm2d(out_channels),)return  blockdef __init__(self, in_channel, out_channel):super(UNet, self).__init__()#Encodeself.conv_encode1 = self.contracting_block(in_channels=in_channel, out_channels=64)self.conv_maxpool1 = torch.nn.MaxPool2d(kernel_size=2)self.conv_encode2 = self.contracting_block(64, 128)self.conv_maxpool2 = torch.nn.MaxPool2d(kernel_size=2)self.conv_encode3 = self.contracting_block(128, 256)self.conv_maxpool3 = torch.nn.MaxPool2d(kernel_size=2)# Bottleneckself.bottleneck = torch.nn.Sequential(torch.nn.Conv2d(kernel_size=3, in_channels=256, out_channels=512),torch.nn.ReLU(),torch.nn.BatchNorm2d(512),torch.nn.Conv2d(kernel_size=3, in_channels=512, out_channels=512),torch.nn.ReLU(),torch.nn.BatchNorm2d(512),torch.nn.ConvTranspose2d(in_channels=512, out_channels=256, kernel_size=3, stride=2, padding=1, output_padding=1))# Decodeself.conv_decode3 = self.expansive_block(512, 256, 128)self.conv_decode2 = self.expansive_block(256, 128, 64)self.final_layer = self.final_block(128, 64, out_channel)def crop_and_concat(self, upsampled, bypass, crop=False):if crop:c = (bypass.size()[2] - upsampled.size()[2]) // 2bypass = F.pad(bypass, (-c, -c, -c, -c))return torch.cat((upsampled, bypass), 1)def forward(self, x):# Encodeencode_block1 = self.conv_encode1(x)encode_pool1 = self.conv_maxpool1(encode_block1)encode_block2 = self.conv_encode2(encode_pool1)encode_pool2 = self.conv_maxpool2(encode_block2)encode_block3 = self.conv_encode3(encode_pool2)encode_pool3 = self.conv_maxpool3(encode_block3)# Bottleneckbottleneck1 = self.bottleneck(encode_pool3)# Decodedecode_block3 = self.crop_and_concat(bottleneck1, encode_block3, crop=True)cat_layer2 = self.conv_decode3(decode_block3)decode_block2 = self.crop_and_concat(cat_layer2, encode_block2, crop=True)cat_layer1 = self.conv_decode2(decode_block2)decode_block1 = self.crop_and_concat(cat_layer1, encode_block1, crop=True)final_layer = self.final_layer(decode_block1)return  final_layer

上面代码中的 UNet 模块代表了 UNet 的整个架构。contraction_block和expansive_block分别用于创建收缩段和膨胀段。该函数crop_and_concat将收缩层的输出与新的扩展层输入相加。

unet = Unet(in_channel=1,out_channel=2)
#out_channel represents number of segments desired
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(unet.parameters(), lr = 0.01, momentum=0.99)
optimizer.zero_grad()
outputs = unet(inputs)
# permute such that number of desired segments would be on 4th dimension
outputs = outputs.permute(0, 2, 3, 1)
m = outputs.shape[0]
# Resizing the outputs and label to caculate pixel wise softmax loss
outputs = outputs.resize(m*width_out*height_out, 2)
labels = labels.resize(m*width_out*height_out)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()

对于该数据集解决标准教程代码：https://www.kaggle.com/hsankesara/unet-image-segmentation