深度学习《语义分割U-Net》_contracting path

论文标题：
U-Net: Convolutional Networks for Biomedical Image Segmentation
虽然主要运用医学领域但是在其他领域效果也很好。
U-Net网络非常简单，前半部分作用是特征提取，后半部分是上采样。在一些文献中也把这样的结构叫做编码器-解码器结构。由于此网络整体结构类似于大写的英文字母U，故得名U-net。
U-Net与其他常见的分割网络有一点非常不同的地方：U-net采用了完全不同的特征融合方式：拼接，U-net采用将特征在channel维度拼接在一起，形成更厚的特征。而FCN融合时使用的对应点相加，并不形成更厚的特征。
U-Net只是一个网络结构的代号而已，只要符合此类特点都可以叫U-Net。
根据论文原文：

我们把U-Net的左边叫做： contracting path
我们把U-Net的右边叫做： expanding path

contracting path：常规的特征提取网络， 一直down sampling的过程，只是提供了四层的特征输出。
expanding path：一直up sampling的过程，堆叠contracting path提供的四层特征，输出一张tensor
语义分割：就是对像素的分类。

直接写代码，我们迁移VGG16完成，contracting path的实现。

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import datasets, layers, optimizers, Sequential, models, applications

baseModel = applications.VGG16(weights="imagenet",include_top=False,input_shape = (512, 512, 3),pooling=None)
baseModel.trainable = False
contractingNet = models.Model(inputs=baseModel.input, outputs=baseModel.get_layer('block4_pool').output)
contractingNet.summary()
f1 = contractingNet.get_layer("block1_conv2").output
f2 = contractingNet.get_layer("block2_conv2").output
f3 = contractingNet.get_layer("block3_conv3").output
f4 = contractingNet.get_layer("block4_conv3").output
1
2
3
4
5
6
7
8
9
10
11
12

expanding path实现：

a = layers.Conv2D(1024, kernel_size=[3, 3], padding="same", activation="relu")(contractingNet.output)
a = layers.Conv2D(512, kernel_size=[1, 1], padding="same", activation="relu")(a)
a = layers.Conv2D(1024, kernel_size=[3, 3], padding="same", activation="relu")(a)

a = layers.UpSampling2D(size = (2,2))(a)
a = layers.concatenate([a,f4], axis = -1)
a = layers.Conv2D(512, kernel_size=[3, 3], padding="same", activation="relu")(a)
a = layers.Conv2D(256, kernel_size=[1, 1], padding="same", activation="relu")(a)
a = layers.Conv2D(512, kernel_size=[3, 3], padding="same", activation="relu")(a)

a = layers.UpSampling2D(size = (2,2))(a)
a = layers.concatenate([a,f3], axis = -1)
a = layers.Conv2D(256, kernel_size=[3, 3], padding="same", activation="relu")(a)
a = layers.Conv2D(128, kernel_size=[1, 1], padding="same", activation="relu")(a)
a = layers.Conv2D(256, kernel_size=[3, 3], padding="same", activation="relu")(a)

a = layers.UpSampling2D(size = (2,2))(a)
a = layers.concatenate([a,f2], axis = -1)
a = layers.Conv2D(128, kernel_size=[3, 3], padding="same", activation="relu")(a)
a = layers.Conv2D(64, kernel_size=[1, 1], padding="same", activation="relu")(a)
a = layers.Conv2D(128, kernel_size=[3, 3], padding="same", activation="relu")(a)

a = layers.UpSampling2D(size = (2,2))(a)
a = layers.concatenate([a,f1], axis = -1)
a = layers.Conv2D(64, kernel_size=[3, 3], padding="same", activation="relu")(a)
a = layers.Conv2D(32, kernel_size=[1, 1], padding="same", activation="relu")(a)
a = layers.Conv2D(64, kernel_size=[3, 3], padding="same", activation="relu")(a)

a = layers.Conv2D(8, kernel_size=[3, 3], padding="same", activation="relu")(a)
outputs = layers.Conv2D(1, kernel_size=[1, 1], padding="same", activation="sigmoid")(a)
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

就这样实现了一个自己的U-Net网络了。
再看看自己写的U-Net训练完的效果图。




论文：http://www.arxiv.org/pdf/1505.04597.pdf
源代码：https://github.com/wuyouyin/tensorflow2-unet
dataset_cell：https://pan.baidu.com/s/1PfOMxGUwJhO-9p8yDsasag 提取码：nqhh
dataset_sidewalk：https://pan.baidu.com/s/1SZ42yGLNf2C9XUNUQkQCbw 提取码：ntyh
dataset_trafficRoad：https://pan.baidu.com/s/1mFFAx7dPcrdkXsbGQomGRg 提取码：20oh