Python深度学习之Unet 语义分割模型(Keras)

前言

最近由于在寻找方向上迷失自我，准备了解更多的计算机视觉任务重的模型。看到语义分割任务重Unet一个有意思的模型，我准备来复现一下它。

一、什么是语义分割

语义分割任务，如下图所示：

简而言之，语义分割任务就是将图片中的不同类别，用不同的颜色标记出来，每一个类别使用一种颜色。常用于医学图像，卫星图像任务。

那如何做到将像素点上色呢？

其实语义分割的输出和图像分类网络类似，图像分类类别数是一个一维的one hot 矩阵。例如：三分类的[0,1,0]。

语义分割任务最后的输出特征图 是一个三维结构，大小与原图类似，通道数就是类别数。 如下图（图片来源于知乎）所示：

其中通道数是类别数，每个通道所标记的像素点，是该类别在图像中的位置，最后通过argmax 取每个通道有用像素 合成一张图像，用不同颜色表示其类别位置。 语义分割任务其实也是分类任务中的一种，他不过是对每一个像素点进行细分，找到每一个像素点所述的类别。 这就是语义分割任务啦~

下面我们来复现 unet 模型

二、Unet

1.基本原理

什么是Unet，它的网络结构如下图所示：

整个网络是一个“U” 的形状，Unet 网络可以分成两部分，上图红色方框中是特征提取部分，和其他卷积神经网络一样，都是通过堆叠卷积提取图像特征，通过池化来压缩特征图。蓝色方框中为图像还原部分（这样称它可能不太专业，大家理解就好），通过上采样和卷积来来将压缩的图像进行还原。特征提取部分可以使用优秀的网络，例如：Resnet50，VGG等。

注意：由于 Resnet50和VGG 网络太大。本文将使用Mobilenet 作为主干特征提取网络。为了方便理解Unet，本文将使用自己搭建的一个mini_unet 去帮祝大家理解。为了方便计算，复现过程会把压缩后的特征图上采样和输入的特征图一样大小。

代码github地址: 一直上不去

先上传到码云： https://gitee.com/Boss-Jian/unet

2.mini_unet

mini_unet 是搭建来帮助大家理解语义分割的网络流程，并不能作为一个优秀的模型完成语义分割任务，来看一下代码的实现：

from keras.layers import Input,Conv2D,Dropout,MaxPooling2D,Concatenate,UpSampling2D
from numpy import pad
from keras.models import Model
def unet_mini(n_classes=21,input_shape=(224,224,3)):

img_input = Input(shape=input_shape)

#------------------------------------------------------
   # #encoder 部分
   #224,224,3 - > 112,112,32
   conv1 = Conv2D(32,(3,3),activation='relu',padding='same')(img_input)
   conv1 = Dropout(0.2)(conv1)
   conv1 = Conv2D(32,(3,3),activation='relu',padding='same')(conv1)
   pool1 = MaxPooling2D((2,2),strides=2)(conv1)

#112,112,32 -> 56,56,64
   conv2 = Conv2D(64,(3,3),activation='relu',padding='same')(pool1)
   conv2 = Dropout(0.2)(conv2)
   conv2 = Conv2D(64,(3,3),activation='relu',padding='same')(conv2)
   pool2 = MaxPooling2D((2,2),strides=2)(conv2)

#56,56,64 -> 56,56,128
   conv3 = Conv2D(128,(3,3),activation='relu',padding='same')(pool2)
   conv3 = Dropout(0.2)(conv3)
   conv3 = Conv2D(128,(3,3),activation='relu',padding='same')(conv3)

#-------------------------------------------------
   # decoder 部分
   #56,56,128 -> 112,112,64
   up1 = UpSampling2D(2)(conv3)
   #112,112,64 -> 112,112,64+128
   up1 = Concatenate(axis=-1)([up1,conv2])
   #  #112,112,192 -> 112,112,64
   conv4  = Conv2D(64,(3,3),activation='relu',padding='same')(up1)
   conv4  = Dropout(0.2)(conv4)
   conv4  = Conv2D(64,(3,3),activation='relu',padding='same')(conv4)

#112,112,64 - >224,224,64
   up2 = UpSampling2D(2)(conv4)
   #224,224,64 -> 224,224,64+32
   up2 = Concatenate(axis=-1)([up2,conv1])
   # 224,224,96 -> 224,224,32
   conv5 =  Conv2D(32,(3,3),activation='relu',padding='same')(up2)
   conv5  = Dropout(0.2)(conv5)
   conv5  = Conv2D(32,(3,3),activation='relu',padding='same')(conv5)

o = Conv2D(n_classes,1,padding='same')(conv5)

return Model(img_input,o,name="unet_mini")

if __name__=="__main__":
   model = unet_mini()
   model.summary()

mini_unet 通过encoder 部分将 224x224x3的图像 变成 112x112x64 的特征图，再通过 上采样方法将特征图放大到 224x224x32。最后通过卷积：

o = Conv2D(n_classes,1,padding='same')(conv5)

将特征图的通道数调节成和类别数一样。

3. Mobilenet_unet

Mobilenet_unet 是使用Mobinet 作为主干特征提取网络，并且加载预训练权重来提升特征提取的能力。decoder 的还原部分和上面一致，下面是Mobilenet_unet 的网络结构：

from keras.models import *
from keras.layers import *
import keras.backend as K
import keras
from tensorflow.python.keras.backend import shape

IMAGE_ORDERING =  "channels_last"# channel last
def relu6(x):
   return K.relu(x, max_value=6)

def _conv_block(inputs, filters, alpha, kernel=(3, 3), strides=(1, 1)):

channel_axis = 1 if IMAGE_ORDERING == 'channels_first' else -1
   filters = int(filters * alpha)
   x = ZeroPadding2D(padding=(1, 1), name='conv1_pad',
                     data_format=IMAGE_ORDERING)(inputs)
   x = Conv2D(filters, kernel, data_format=IMAGE_ORDERING,
              padding='valid',
              use_bias=False,
              strides=strides,
              name='conv1')(x)
   x = BatchNormalization(axis=channel_axis, name='conv1_bn')(x)
   return Activation(relu6, name='conv1_relu')(x)

def _depthwise_conv_block(inputs, pointwise_conv_filters, alpha,
                         depth_multiplier=1, strides=(1, 1), block_id=1):

channel_axis = 1 if IMAGE_ORDERING == 'channels_first' else -1
   pointwise_conv_filters = int(pointwise_conv_filters * alpha)

x = ZeroPadding2D((1, 1), data_format=IMAGE_ORDERING,
                     name='conv_pad_％d' ％ block_id)(inputs)
   x = DepthwiseConv2D((3, 3), data_format=IMAGE_ORDERING,
                       padding='valid',
                       depth_multiplier=depth_multiplier,
                       strides=strides,
                       use_bias=False,
                       name='conv_dw_％d' ％ block_id)(x)
   x = BatchNormalization(
       axis=channel_axis, name='conv_dw_％d_bn' ％ block_id)(x)
   x = Activation(relu6, name='conv_dw_％d_relu' ％ block_id)(x)

x = Conv2D(pointwise_conv_filters, (1, 1), data_format=IMAGE_ORDERING,
              padding='same',
              use_bias=False,
              strides=(1, 1),
              name='conv_pw_％d' ％ block_id)(x)
   x = BatchNormalization(axis=channel_axis,
                          name='conv_pw_％d_bn' ％ block_id)(x)
   return Activation(relu6, name='conv_pw_％d_relu' ％ block_id)(x)

def get_mobilnet_eocoder(input_shape=(224,224,3),weights_path=""):

# 必须是32 的倍数
   assert input_shape[0] ％ 32 == 0
   assert input_shape[1] ％ 32 == 0

alpha = 1.0
   depth_multiplier = 1

img_input = Input(shape=input_shape)
   #(None, 224, 224, 3) ->(None, 112, 112, 64)
   x = _conv_block(img_input, 32, alpha, strides=(2, 2))
   x = _depthwise_conv_block(x, 64, alpha, depth_multiplier, block_id=1)
   f1 = x

#(None, 112, 112, 64) -> (None, 56, 56, 128)
   x = _depthwise_conv_block(x, 128, alpha, depth_multiplier,
                             strides=(2, 2), block_id=2)
   x = _depthwise_conv_block(x, 128, alpha, depth_multiplier, block_id=3)
   f2 = x
  #(None, 56, 56, 128) -> (None, 28, 28, 256)
   x = _depthwise_conv_block(x, 256, alpha, depth_multiplier,
                             strides=(2, 2), block_id=4)
   x = _depthwise_conv_block(x, 256, alpha, depth_multiplier, block_id=5)
   f3 = x
   # (None, 28, 28, 256) ->  (None, 14, 14, 512)
   x = _depthwise_conv_block(x, 512, alpha, depth_multiplier,
                             strides=(2, 2), block_id=6)
   x = _depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=7)
   x = _depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=8)
   x = _depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=9)
   x = _depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=10)
   x = _depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=11)
   f4 = x
   # (None, 14, 14, 512) -> (None, 7, 7, 1024)
   x = _depthwise_conv_block(x, 1024, alpha, depth_multiplier,
                             strides=(2, 2), block_id=12)
   x = _depthwise_conv_block(x, 1024, alpha, depth_multiplier, block_id=13)
   f5 = x
   # 加载预训练模型
   if weights_path!="":
       Model(img_input, x).load_weights(weights_path, by_name=True, skip_mismatch=True)
   # f1: (None, 112, 112, 64)
   # f2: (None, 56, 56, 128)
   # f3: (None, 28, 28, 256)
   # f4: (None, 14, 14, 512)
   # f5: (None, 7, 7, 1024)
   return img_input, [f1, f2, f3, f4, f5]

def mobilenet_unet(num_classes=2,input_shape=(224,224,3)):

#encoder
   img_input,levels = get_mobilnet_eocoder(input_shape=input_shape,weights_path="model_data\mobilenet_1_0_224_tf_no_top.h5")

[f1, f2, f3, f4, f5] = levels

# f1: (None, 112, 112, 64)
   # f2: (None, 56, 56, 128)
   # f3: (None, 28, 28, 256)
   # f4: (None, 14, 14, 512)
   # f5: (None, 7, 7, 1024)

#decoder
   #(None, 14, 14, 512) - > (None, 14, 14, 512)
   o = f4
   o = ZeroPadding2D()(o)
   o = Conv2D(512, (3, 3), padding='valid' , activation='relu' , data_format=IMAGE_ORDERING)(o)
   o = BatchNormalization()(o)

#(None, 14, 14, 512) ->(None,28,28,256)
   o = UpSampling2D(2)(o)
   o = Concatenate(axis=-1)([o,f3])
   o = ZeroPadding2D()(o)
   o = Conv2D(256, (3, 3), padding='valid' , activation='relu' , data_format=IMAGE_ORDERING)(o)
   o = BatchNormalization()(o)
   # None,28,28,256)->(None,56,56,128)
   o = UpSampling2D(2)(o)
   o = Concatenate(axis=-1)([o,f2])
   o = ZeroPadding2D()(o)
   o = Conv2D(128, (3, 3), padding='valid' , activation='relu' , data_format=IMAGE_ORDERING)(o)
   o = BatchNormalization()(o)
   #(None,56,56,128) ->(None,112,112,64)
   o = UpSampling2D(2)(o)
   o = Concatenate(axis=-1)([o,f1])
   o = ZeroPadding2D()(o)
   o = Conv2D(128, (3, 3), padding='valid' , activation='relu' , data_format=IMAGE_ORDERING)(o)
   o = BatchNormalization()(o)
   #(None,112,112,64) -> (None,112,112,num_classes)

# 再上采样 让输入和出处图片大小一致
   o = UpSampling2D(2)(o)
   o = ZeroPadding2D()(o)
   o = Conv2D(64, (3, 3), padding='valid' , activation='relu' , data_format=IMAGE_ORDERING)(o)
   o = BatchNormalization()(o)

o = Conv2D(num_classes, (3, 3), padding='same',
              data_format=IMAGE_ORDERING)(o)

return Model(img_input,o)

if __name__=="__main__":
   mobilenet_unet(input_shape=(512,512,3)).summary()

特征图的大小变化，以及代码含义都已经注释在代码里了。大家仔细阅读吧

4.数据加载部分

import math
import os
from random import shuffle

import cv2
import keras
import numpy as np
from PIL import Image
#-------------------------------
# 将图片转换为 rgb
#------------------------------
def cvtColor(image):
   if len(np.shape(image)) == 3 and np.shape(image)[2] == 3:
       return image
   else:
       image = image.convert('RGB')
       return image
#-------------------------------
# 图片归一化 0~1
#------------------------------
def preprocess_input(image):
   image = image / 127.5 - 1
   return image
#---------------------------------------------------
#   对输入图像进行resize
#---------------------------------------------------
def resize_image(image, size):
   iw, ih  = image.size
   w, h    = size

scale   = min(w/iw, h/ih)
   nw      = int(iw*scale)
   nh      = int(ih*scale)

image   = image.resize((nw,nh), Image.BICUBIC)
   new_image = Image.new('RGB', size, (128,128,128))
   new_image.paste(image, ((w-nw)//2, (h-nh)//2))

return new_image, nw, nh

class UnetDataset(keras.utils.Sequence):
   def __init__(self, annotation_lines, input_shape, batch_size, num_classes, train, dataset_path):
       self.annotation_lines   = annotation_lines
       self.length             = len(self.annotation_lines)
       self.input_shape        = input_shape
       self.batch_size         = batch_size
       self.num_classes        = num_classes
       self.train              = train
       self.dataset_path       = dataset_path

def __len__(self):
       return math.ceil(len(self.annotation_lines) / float(self.batch_size))

def __getitem__(self, index):
       #图片和标签、
       images  = []
       targets = []
       # 读取一个batchsize
       for i in range(index*self.batch_size,(index+1)*self.batch_size):
           #判断 i 越界情况
           i = i％self.length
           name = self.annotation_lines[i].split()[0]
           # 从路径中读取图像 jpg 表示图片，png 表示标签
           jpg = Image.open(os.path.join(os.path.join(self.dataset_path,'Images'),name+'.png'))
           png = Image.open(os.path.join(os.path.join(self.dataset_path,'Labels'),name+'.png'))

#-------------------
           # 数据增强  和 归一化
           #-------------------
           jpg,png = self.get_random_data(jpg,png,self.input_shape,random=self.train)
           jpg = preprocess_input(np.array(jpg,np.float64))
           png = np.array(png)

#-----------------------------------
           # 医学图像中 描绘出的是细胞边缘
           #  将小于 127.5的像素点 作为目标 像素点
           #------------------------------------

seg_labels = np.zeros_like(png)
           seg_labels[png<=127.5] = 1
           #--------------------------------
           # 转化为 one hot 标签
           # -------------------------
           seg_labels  = np.eye(self.num_classes + 1)[seg_labels.reshape([-1])]
           seg_labels  = seg_labels.reshape((int(self.input_shape[0]), int(self.input_shape[1]), self.num_classes + 1))

images.append(jpg)
           targets.append(seg_labels)

images  = np.array(images)
       targets = np.array(targets)
       return images, targets

def rand(self, a=0, b=1):
       return np.random.rand() * (b - a) + a

def get_random_data(self, image, label, input_shape, jitter=.3, hue=.1, sat=1.5, val=1.5, random=True):
       image = cvtColor(image)
       label = Image.fromarray(np.array(label))
       h, w = input_shape

if not random:
           iw, ih  = image.size
           scale   = min(w/iw, h/ih)
           nw      = int(iw*scale)
           nh      = int(ih*scale)

image       = image.resize((nw,nh), Image.BICUBIC)
           new_image   = Image.new('RGB', [w, h], (128,128,128))
           new_image.paste(image, ((w-nw)//2, (h-nh)//2))

label       = label.resize((nw,nh), Image.NEAREST)
           new_label   = Image.new('L', [w, h], (0))
           new_label.paste(label, ((w-nw)//2, (h-nh)//2))
           return new_image, new_label

# resize image
       rand_jit1 = self.rand(1-jitter,1+jitter)
       rand_jit2 = self.rand(1-jitter,1+jitter)
       new_ar = w/h * rand_jit1/rand_jit2

scale = self.rand(0.25, 2)
       if new_ar < 1:
           nh = int(scale*h)
           nw = int(nh*new_ar)
       else:
           nw = int(scale*w)
           nh = int(nw/new_ar)

image = image.resize((nw,nh), Image.BICUBIC)
       label = label.resize((nw,nh), Image.NEAREST)

flip = self.rand()<.5
       if flip:
           image = image.transpose(Image.FLIP_LEFT_RIGHT)
           label = label.transpose(Image.FLIP_LEFT_RIGHT)

# place image
       dx = int(self.rand(0, w-nw))
       dy = int(self.rand(0, h-nh))
       new_image = Image.new('RGB', (w,h), (128,128,128))
       new_label = Image.new('L', (w,h), (0))
       new_image.paste(image, (dx, dy))
       new_label.paste(label, (dx, dy))
       image = new_image
       label = new_label

# distort image
       hue = self.rand(-hue, hue)
       sat = self.rand(1, sat) if self.rand()<.5 else 1/self.rand(1, sat)
       val = self.rand(1, val) if self.rand()<.5 else 1/self.rand(1, val)
       x = cv2.cvtColor(np.array(image,np.float32)/255, cv2.COLOR_RGB2HSV)
       x[..., 0] += hue*360
       x[..., 0][x[..., 0]>1] -= 1
       x[..., 0][x[..., 0]<0] += 1
       x[..., 1] *= sat
       x[..., 2] *= val
       x[x[:,:, 0]>360, 0] = 360
       x[:, :, 1:][x[:, :, 1:]>1] = 1
       x[x<0] = 0
       image_data = cv2.cvtColor(x, cv2.COLOR_HSV2RGB)*255
       return image_data,label

def on_epoch_begin(self):
       shuffle(self.annotation_lines)

训练过程代码：

import numpy as np
from  tensorflow.python.keras.callbacks import EarlyStopping, ModelCheckpoint, TensorBoard
from keras.optimizers import Adam
import os
from unet_mini import unet_mini
from mobilnet_unet import mobilenet_unet
from callbacks import ExponentDecayScheduler,LossHistory
from keras import backend as K
from keras import backend
from data_loader import UnetDataset
#--------------------------------------
# 交叉熵损失函数 cls_weights 类别的权重
#-------------------------------------
def CE(cls_weights):
   cls_weights = np.reshape(cls_weights, [1, 1, 1, -1])
   def _CE(y_true, y_pred):
       y_pred = K.clip(y_pred, K.epsilon(), 1.0 - K.epsilon())

CE_loss = - y_true[...,:-1] * K.log(y_pred) * cls_weights
       CE_loss = K.mean(K.sum(CE_loss, axis = -1))
       # dice_loss = tf.Print(CE_loss, [CE_loss])
       return CE_loss
   return _CE
def f_score(beta=1, smooth = 1e-5, threhold = 0.5):
   def _f_score(y_true, y_pred):
       y_pred = backend.greater(y_pred, threhold)
       y_pred = backend.cast(y_pred, backend.floatx())

tp = backend.sum(y_true[...,:-1] * y_pred, axis=[0,1,2])
       fp = backend.sum(y_pred         , axis=[0,1,2]) - tp
       fn = backend.sum(y_true[...,:-1], axis=[0,1,2]) - tp

score = ((1 + beta ** 2) * tp + smooth) \
               / ((1 + beta ** 2) * tp + beta ** 2 * fn + fp + smooth)
       return score
   return _f_score

def train():
   #-------------------------
   # 细胞图像 分为细胞壁 和其他
   # 初始化 参数
   #-------------------------
   num_classes  = 2

input_shape = (512,512,3)
   # 从第几个epoch 继续训练

batch_size = 4

learn_rate  = 1e-4

start_epoch = 0
   end_epoch = 100
   num_workers = 4

dataset_path = 'Medical_Datasets'

model = mobilenet_unet(num_classes,input_shape=input_shape)

model.summary()

# 读取数据图片的路劲
   with open(os.path.join(dataset_path, "ImageSets/Segmentation/train.txt"),"r") as f:
       train_lines = f.readlines()

logging         = TensorBoard(log_dir = 'logs/')
   checkpoint      = ModelCheckpoint('logs/ep{epoch:03d}-loss{loss:.3f}.h5',
                       monitor = 'loss', save_weights_only = True, save_best_only = False, period = 1)
   reduce_lr       = ExponentDecayScheduler(decay_rate = 0.96, verbose = 1)
   early_stopping  = EarlyStopping(monitor='loss', min_delta=0, patience=10, verbose=1)
   loss_history    = LossHistory('logs/', val_loss_flag = False)

epoch_step      = len(train_lines) // batch_size
   cls_weights     = np.ones([num_classes], np.float32)
   loss = CE(cls_weights)
   model.compile(loss = loss,
               optimizer = Adam(lr=learn_rate),
               metrics = [f_score()])

train_dataloader    = UnetDataset(train_lines, input_shape[:2], batch_size, num_classes, True, dataset_path)

print('Train on {} samples, with batch size {}.'.format(len(train_lines), batch_size))
   model.fit_generator(
           generator           = train_dataloader,
           steps_per_epoch     = epoch_step,
           epochs              = end_epoch,
           initial_epoch       = start_epoch,
           # use_multiprocessing = True if num_workers > 1 else False,
           workers             = num_workers,
           callbacks           = [logging, checkpoint, early_stopping,reduce_lr,loss_history]
       )

if __name__=="__main__":
   train()

最后的预测结果:

完整的代大家感兴趣可以去github下载下来再看，代码比较多，全部贴出来博客显得太长了。

这就是简单的语义分割任务啦。

参考

https://github.com/bubbliiiing/unet-keras

https://github.com/divamgupta/image-segmentation-keras 

来源：https://blog.csdn.net/qq_38676487/article/details/121903186