【613】U-Net 相关

参考：keras遥感图像Unet语义分割(支持多波段&多类)

参考：憨批的语义分割重制版5——Keras 搭建自己的Unet语义分割平台

参考：U-net源码分析（Keras版本）

参考：U-Net 源码 McDelfino / unet 【国内反而不稳定】

参考：U-Net 源码 alexbnlee/unet【GitHub】

---

　　通过 keras API 模型比较容易，有几种实现方式

• 从头到尾实现，这个最直观简单

• 通过 for 循环实现，有些重复的部分

• 通过写函数构建不同的模块

　　目前看来整个模型实现很容易，但是最开始接触的时候反而是看了很多资料都不太懂，归其原因是对于整体深度学习的理解还不是很到位。目前看来很多模型就是在 U-Net 基础上进行一些改进，包括增加 BatchNormalization 层，或者增加 ResNet 层，或者添加 Attention 层，怎么说呢，慢慢测试吧。

　　U-Net 模型实现：

?

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

import numpy as np import os import skimage.io as io import skimage.transform as trans import numpy as np from keras.models import * from keras.layers import * from keras.optimizers import * from keras.callbacks import ModelCheckpoint, LearningRateScheduler from keras import backend as keras     def unet(pretrained_weights = None,input_size = (256,256,1)):     inputs = Input(input_size)     conv1 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(inputs)     conv1 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv1)     pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)     conv2 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool1)     conv2 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv2)     pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)     conv3 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool2)     conv3 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv3)     pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)     conv4 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool3)     conv4 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv4)     drop4 = Dropout(0.5)(conv4)     pool4 = MaxPooling2D(pool_size=(2, 2))(drop4)       conv5 = Conv2D(1024, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool4)     conv5 = Conv2D(1024, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv5)     drop5 = Dropout(0.5)(conv5)       up6 = Conv2D(512, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(drop5))     merge6 = concatenate([drop4,up6], axis = 3)     conv6 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge6)     conv6 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv6)       up7 = Conv2D(256, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv6))     merge7 = concatenate([conv3,up7], axis = 3)     conv7 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge7)     conv7 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv7)       up8 = Conv2D(128, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv7))     merge8 = concatenate([conv2,up8], axis = 3)     conv8 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge8)     conv8 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv8)       up9 = Conv2D(64, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv8))     merge9 = concatenate([conv1,up9], axis = 3)     conv9 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge9)     conv9 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv9)     conv9 = Conv2D(2, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv9)     conv10 = Conv2D(1, 1, activation = 'sigmoid')(conv9)       model = Model(input = inputs, output = conv10)       model.compile(optimizer = Adam(lr = 1e-4), loss = 'binary_crossentropy', metrics = ['accuracy'])           #model.summary()       if(pretrained_weights):         model.load_weights(pretrained_weights)       return model