深度學習-Tensorflow2.2-圖像處理{10}-圖像語義分割-23

圖像語義分割簡介

圖像語義分割網絡結構-FCN

上采樣

代碼實現

import tensorflow as tf
import matplotlib.pyplot as plt
%matplotlib inline
import numpy as np
import glob
import os

# 顯存自適應分配
gpus = tf.config.experimental.list_physical_devices(device_type='GPU')
for gpu in gpus:
    tf.config.experimental.set_memory_growth(gpu,True)

gpu_ok = tf.test.is_gpu_available()
print("tf version:", tf.__version__)
print("use GPU", gpu_ok) # 判斷是否使用gpu進行訓練

例子

os.listdir("F:/py/ziliao/數據集/圖片定位與分割數據集/annotations/trimaps")[-5:]

# 讀取圖片
img = tf.io.read_file(r"F:/py/ziliao/數據集/圖片定位與分割數據集/annotations/trimaps/Abyssinian_2.png")

# 讀取圖片
img2 = tf.io.read_file(r"F:/py/ziliao/數據集/圖片定位與分割數據集/images/Abyssinian_2.jpg")

# 解碼
img = tf.image.decode_png(img)
img2 = tf.image.decode_jpeg(img2)

# 查看大小
img.shape

# 移除的所有大小為1的維度具有相同類型的張量
img = tf.squeeze(img)
img.shape

# 繪圖
plt.imshow(img)

img.numpy()

np.unique(img.numpy()) # 我們能看見改圖中只有3種像素 分別對應背景 邊框 身體

plt.imshow(img2)

完整代碼

 # 讀取所有圖片路徑
images = glob.glob(r"F:/py/ziliao/數據集/圖片定位與分割數據集/images/*.jpg")

# 讀取對應的目標圖像
anno = glob.glob(r"F:/py/ziliao/數據集/圖片定位與分割數據集/annotations/trimaps/*.png")

# 對圖片進行隨機
np.random.seed(2020) #設置隨機數種子生成的隨機數將會是同一個
index = np.random.permutation(len(images))

images = np.array(images)[index]
anno = np.array(anno)[index]

# 創建數據集
dataset = tf.data.Dataset.from_tensor_slices((images,anno))

# 划分數據集
test_count = int(len(images)*0.2)
train_count = len(images)-test_count
test_count,train_count

# 分割數據集
data_train =dataset.skip(test_count) # 跳過test_count 的數據作為訓練數據
data_test = dataset.take(test_count) # 取出test_count的數據作為測試集

# 創建 jpg格式的解碼函數
def read_jpg(path):
    img = tf.io.read_file(path)
    img = tf.image.decode_jpeg(img,channels=3)
    return img

# 創建png的解碼函數
def read_png(path):
    img = tf.io.read_file(path)
    img = tf.image.decode_png(img,channels=1)
    return img

# 歸一化
def normal_img(input_images,input_anno):
    input_images = tf.cast(input_images,tf.float32) # 改變數據類型為float32
    input_images/127.5 - 1  # 歸一化到 -1 到 1 之間
    
    input_anno -= 1  # 因為目標圖像取值范圍是 1，2，3 我們 -1 修改為 0，1，2
    return input_images,input_anno

# 加載函數
def load_images(input_images_path,input_anno_path):
    input_image = read_jpg(input_images_path)
    input_anno = read_png(input_anno_path)
    input_image = tf.image.resize(input_image,(224,224))
    input_anno = tf.image.resize(input_anno,(224,224))
    return normal_img(input_image,input_anno)

# 使用加載函數
data_train = data_train.map(load_images,
                            num_parallel_calls=tf.data.experimental.AUTOTUNE)

data_test = data_test.map(load_images,
                            num_parallel_calls=tf.data.experimental.AUTOTUNE)

BATCH_SIZE = 16

# 對數據集進行亂序
data_train = data_train.repeat().shuffle(5912).batch(BATCH_SIZE)
data_test = data_test.batch(BATCH_SIZE)

for img,anno in data_train.take(1):
    plt.subplot(1,2,1) # 1行2列
    plt.imshow(tf.keras.preprocessing.image.array_to_img(img[0]))
    plt.subplot(1,2,2) # 1行2列
    plt.imshow(tf.keras.preprocessing.image.array_to_img(anno[0]))

# 使用預訓練網絡
conv_base = tf.keras.applications.VGG16(weights="imagenet", # 使用該模型image權重
                                        input_shape=(224,224,3), 
                                        include_top = False)

conv_base.summary()

# 例： 獲得某一層的輸出 如 獲取 block5_conv3 的層的輸出
conv_base.get_layer("block5_conv3").output

# 例： 從預訓練網絡中創建子模型
sub_model = tf.keras.models.Model(inputs = conv_base.input,
                                  outputs = conv_base.get_layer("block5_conv3").output
                                 )

FCN跳階-獲取模型中間層的輸出

layer_names = [
    "block5_conv3", # 14x14
    "block4_conv3", # 28x28
    "block3_conv3", # 56x56
    "block5_pool"
] # 獲取輸出層的名字

# 創建特征提取模型
layers_output = [conv_base.get_layer(layer_name).output for layer_name in layer_names]

# 從預訓練網絡中創建子模型
multi_out_model = tf.keras.models.Model(inputs = conv_base.input,
                                        outputs = layers_output
                                 )

multi_out_model.trainable = False # 禁止訓練

#構建模型
inputs = tf.keras.layers.Input(shape=(224,224,3))
out_block5_conv3,out_block4_conv3,out_block3_conv3,out = multi_out_model(inputs)

反卷積 上采樣

#輸出512個卷積核 窗口大小3*3 圖片變大2倍填充方式same保證和原有圖像大小一樣 激活relu
x1 = tf.keras.layers.Conv2DTranspose(512,3,
                                     strides=2,
                                     padding="same",
                                     activation="relu")(out) # 調用out層

x1 = tf.keras.layers.Conv2D(512,3,
                            padding="same",
                            activation="relu")(x1) # 在增加一層卷積

x2 = tf.add(x1,out_block5_conv3) # 層相加
x2.shape

x2 = tf.keras.layers.Conv2DTranspose(512,3,strides=2,padding="same",activation="relu")(x2)

x2 = tf.keras.layers.Conv2D(512,3,padding="same",activation="relu")(x2) # 在增加一層卷積

x3 = tf.add(x2,out_block4_conv3) # 層相加
x3.shape

x3 = tf.keras.layers.Conv2DTranspose(256,3,strides=2,padding="same",activation="relu")(x3)

x3 = tf.keras.layers.Conv2D(256,3,padding="same",activation="relu")(x3) # 在增加一層卷積進一步提取特征

x4 = tf.add(x3,out_block3_conv3)
x4.shape

x5 = tf.keras.layers.Conv2DTranspose(128,3,strides=2,padding="same",activation="relu")(x4)

x5 = tf.keras.layers.Conv2D(128,3,padding="same",activation="relu")(x5) # 在增加一層卷積進一步提取特征

prediction = tf.keras.layers.Conv2DTranspose(3, # 因為目標圖像取值是3個，所以我們輸出為3
                                             3, 
                                             strides=2,
                                             padding="same",
                                             activation="softmax")(x5) # 上采樣

model = tf.keras.models.Model(
    inputs=inputs,
    outputs=prediction
)

# 模型編譯
model.compile(
    optimizer="adam",
    loss = "sparse_categorical_crossentropy",
    metrics=["acc"]
)

# 訓練模型
history = model.fit(data_train,
                    epochs=5,
                    steps_per_epoch=train_count//BATCH_SIZE,
                    validation_data=data_test,
                    validation_steps=test_count//BATCH_SIZE,
    
)

loss = history.history["loss"]
val_loss = history.history["val_loss"]

epochs = range(15)

plt.figure()
plt.plot(epochs,loss,"r",label="Trainning loss")
plt.plot(epochs,val_loss,"bo",label="Validation loss")
plt.title("Training and Validation Loss")
plt.xlabel("epochs")
plt.ylabel("Loss Value")

plt.legend()
plt.show()

num = 3

for image, mask in data_test.take(1): # 從test數據取出一個batch
    pred_mask = model.predict(image) # model.predict(image) 對圖片進行預測
    pred_mask = tf.argmax(pred_mask, axis=-1)  # 取出預測最大值
    pred_mask = pred_mask[..., tf.newaxis] # 維度擴展 取前面所有維度
    
    plt.figure(figsize=(10, 10))
    for i in range(num):
        plt.subplot(num, 3, i*num+1)
        plt.imshow(tf.keras.preprocessing.image.array_to_img(image[i])) # 原圖
        plt.subplot(num, 3, i*num+2)
        plt.imshow(tf.keras.preprocessing.image.array_to_img(mask[i])) # 真實分割圖
        plt.subplot(num, 3, i*num+3)
        plt.imshow(tf.keras.preprocessing.image.array_to_img(pred_mask[i])) # 預測分割圖