Keras:Unet網路實現多類語義分割方式

阿新 • • 發佈：2020-06-11

1 介紹

U-Net最初是用來對醫學影象的語義分割，後來也有人將其應用於其他領域。但大多還是用來進行二分類，即將原始影象分成兩個灰度級或者色度，依次找到影象中感興趣的目標部分。

本文主要利用U-Net網路結構實現了多類的語義分割，並展示了部分測試效果，希望對你有用！

2 原始碼

（1）訓練模型

from __future__ import print_function
import os
import datetime
import numpy as np
from keras.models import Model
from keras.layers import Input,concatenate,Conv2D,MaxPooling2D,Conv2DTranspose,AveragePooling2D,Dropout,\
 BatchNormalization
from keras.optimizers import Adam
from keras.layers.convolutional import UpSampling2D,Conv2D
from keras.callbacks import ModelCheckpoint
from keras import backend as K
from keras.layers.advanced_activations import LeakyReLU,ReLU
import cv2
 
PIXEL = 512 #set your image size
BATCH_SIZE = 5
lr = 0.001
EPOCH = 100
X_CHANNEL = 3 # training images channel
Y_CHANNEL = 1 # label iamges channel
X_NUM = 422 # your traning data number
 
pathX = 'I:\\Pascal VOC Dataset\\train1\\images\\' #change your file path
pathY = 'I:\\Pascal VOC Dataset\\train1\\SegmentationObject\\' #change your file path
 
#data processing
def generator(pathX,pathY,BATCH_SIZE):
 while 1:
  X_train_files = os.listdir(pathX)
  Y_train_files = os.listdir(pathY)
  a = (np.arange(1,X_NUM))
  X = []
  Y = []
  for i in range(BATCH_SIZE):
   index = np.random.choice(a)
   # print(index)
   img = cv2.imread(pathX + X_train_files[index],1)
   img = np.array(img).reshape(PIXEL,PIXEL,X_CHANNEL)
   X.append(img)
   img1 = cv2.imread(pathY + Y_train_files[index],1)
   img1 = np.array(img1).reshape(PIXEL,Y_CHANNEL)
   Y.append(img1)
 
  X = np.array(X)
  Y = np.array(Y)
  yield X,Y
 
 #creat unet network
inputs = Input((PIXEL,3))
conv1 = Conv2D(8,3,activation='relu',padding='same',kernel_initializer='he_normal')(inputs)
pool1 = AveragePooling2D(pool_size=(2,2))(conv1) # 16
 
conv2 = BatchNormalization(momentum=0.99)(pool1)
conv2 = Conv2D(64,kernel_initializer='he_normal')(conv2)
conv2 = BatchNormalization(momentum=0.99)(conv2)
conv2 = Conv2D(64,1,kernel_initializer='he_normal')(conv2)
conv2 = Dropout(0.02)(conv2)
pool2 = AveragePooling2D(pool_size=(2,2))(conv2) # 8
 
conv3 = BatchNormalization(momentum=0.99)(pool2)
conv3 = Conv2D(128,kernel_initializer='he_normal')(conv3)
conv3 = BatchNormalization(momentum=0.99)(conv3)
conv3 = Conv2D(128,kernel_initializer='he_normal')(conv3)
conv3 = Dropout(0.02)(conv3)
pool3 = AveragePooling2D(pool_size=(2,2))(conv3) # 4
 
conv4 = BatchNormalization(momentum=0.99)(pool3)
conv4 = Conv2D(256,kernel_initializer='he_normal')(conv4)
conv4 = BatchNormalization(momentum=0.99)(conv4)
conv4 = Conv2D(256,kernel_initializer='he_normal')(conv4)
conv4 = Dropout(0.02)(conv4)
pool4 = AveragePooling2D(pool_size=(2,2))(conv4)
 
conv5 = BatchNormalization(momentum=0.99)(pool4)
conv5 = Conv2D(512,kernel_initializer='he_normal')(conv5)
conv5 = BatchNormalization(momentum=0.99)(conv5)
conv5 = Conv2D(512,kernel_initializer='he_normal')(conv5)
conv5 = Dropout(0.02)(conv5)
pool4 = AveragePooling2D(pool_size=(2,2))(conv4)
# conv5 = Conv2D(35,kernel_initializer='he_normal')(conv4)
# drop4 = Dropout(0.02)(conv5)
pool4 = AveragePooling2D(pool_size=(2,2))(pool3) # 2
pool5 = AveragePooling2D(pool_size=(2,2))(pool4) # 1
 
conv6 = BatchNormalization(momentum=0.99)(pool5)
conv6 = Conv2D(256,kernel_initializer='he_normal')(conv6)
 
conv7 = Conv2D(256,kernel_initializer='he_normal')(conv6)
up7 = (UpSampling2D(size=(2,2))(conv7)) # 2
conv7 = Conv2D(256,kernel_initializer='he_normal')(up7)
merge7 = concatenate([pool4,conv7],axis=3)
 
conv8 = Conv2D(128,kernel_initializer='he_normal')(merge7)
up8 = (UpSampling2D(size=(2,2))(conv8)) # 4
conv8 = Conv2D(128,kernel_initializer='he_normal')(up8)
merge8 = concatenate([pool3,conv8],axis=3)
 
conv9 = Conv2D(64,kernel_initializer='he_normal')(merge8)
up9 = (UpSampling2D(size=(2,2))(conv9)) # 8
conv9 = Conv2D(64,kernel_initializer='he_normal')(up9)
merge9 = concatenate([pool2,conv9],axis=3)
 
conv10 = Conv2D(32,kernel_initializer='he_normal')(merge9)
up10 = (UpSampling2D(size=(2,2))(conv10)) # 16
conv10 = Conv2D(32,kernel_initializer='he_normal')(up10)
 
conv11 = Conv2D(16,kernel_initializer='he_normal')(conv10)
up11 = (UpSampling2D(size=(2,2))(conv11)) # 32
conv11 = Conv2D(8,kernel_initializer='he_normal')(up11)
 
# conv12 = Conv2D(3,kernel_initializer='he_normal')(conv11)
conv12 = Conv2D(3,kernel_initializer='he_normal')(conv11)
 
model = Model(input=inputs,output=conv12)
print(model.summary())
model.compile(optimizer=Adam(lr=1e-3),loss='mse',metrics=['accuracy'])
 
history = model.fit_generator(generator(pathX,BATCH_SIZE),steps_per_epoch=600,nb_epoch=EPOCH)
end_time = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
 
 #save your training model
model.save(r'V1_828.h5')
 
#save your loss data
mse = np.array((history.history['loss']))
np.save(r'V1_828.npy',mse)

（2）測試模型

from keras.models import load_model
import numpy as np
import matplotlib.pyplot as plt
import os
import cv2
 
model = load_model('V1_828.h5')
test_images_path = 'I:\\Pascal VOC Dataset\\test\\test_images\\'
test_gt_path = 'I:\\Pascal VOC Dataset\\test\\SegmentationObject\\'
pre_path = 'I:\\Pascal VOC Dataset\\test\\pre\\'
 
X = []
for info in os.listdir(test_images_path):
 A = cv2.imread(test_images_path + info)
 X.append(A)
 # i += 1
X = np.array(X)
print(X.shape)
Y = model.predict(X)
 
groudtruth = []
for info in os.listdir(test_gt_path):
 A = cv2.imread(test_gt_path + info)
 groudtruth.append(A)
groudtruth = np.array(groudtruth)
 
i = 0
for info in os.listdir(test_images_path):
 cv2.imwrite(pre_path + info,Y[i])
 i += 1
 
a = range(10)
n = np.random.choice(a)
cv2.imwrite('prediction.png',Y[n])
cv2.imwrite('groudtruth.png',groudtruth[n])
fig,axs = plt.subplots(1,3)
# cnt = 1
# for j in range(1):
axs[0].imshow(np.abs(X[n]))
axs[0].axis('off')
axs[1].imshow(np.abs(Y[n]))
axs[1].axis('off')
axs[2].imshow(np.abs(groudtruth[n]))
axs[2].axis('off')
 # cnt += 1
fig.savefig("imagestest.png")
plt.close()

3 效果展示

說明：從左到右依次是預測影象，真實影象，標註影象。可以看出，對於部分資料的分割效果還有待改進，主要原因還是資料集相對複雜，模型難於找到其中的規律。

Keras:Unet網路實現多類語義分割方式