如何从 output 分类器创建分割掩码？

Question

I am a newer in ML and I am trying to apply image segmentation on my gray scale tif images. The images has areas with the value NaN which represents the sea, and areas with value from 0 to 2 which represents land. I create some true masks for training. The masks have area with NaN representing the sea, 0 representing the land, and 1 representing the clouds. I would like to create a segmentation mask which has 3 classes representing the sea, the land, and the clouds.

I refer to TensorFlow tutorial and Google Colab tutorial and have the code below. The output classifier do show somethings but the segmentation mask became 0 overall. Please help and thank you.

from glob import glob
from PIL import Image
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
from tensorflow.python.keras import layers
from tensorflow.python.keras import losses
from tensorflow.python.keras import models

#load images
img = sorted(glob('/content/drive/My Drive/train_sub_5/*.tif'))
mask = sorted(glob('/content/drive/My Drive/train_mask_sub_5/*.tif'))

#split into train and test dataset
img, img_val, mask, mask_val = train_test_split(img, mask, test_size=0.2, random_state=42)

#read images as array and make their shape (512, 512, 1)
train_image = []
for m in img[:]:
    image= Image.open(m)
    img_arr= np.nan_to_num(np.array(image), nan=0)
    stacked_img= np.stack((img_arr,)*1, axis=-1)
    train_image.append(stacked_img)

train_mask = []
for n in mask[:]:
    image_mask= Image.open(n)
    mask_arr= np.nan_to_num(np.array(image_mask), nan=2)
    stacked_mask = np.stack((mask_arr,)*1, axis=-1)
    train_mask.append(stacked_mask)

test_img = []
for o in img_val[:]:
    image= Image.open(o)
    img_arr = np.nan_to_num(np.array(image), nan=0)
    stacked_img = np.stack((img_arr,)*1, axis=-1)
    test_img.append(stacked_img)

test_mask = []
for p in mask_val[:]:
    image_mask= Image.open(p)
    mask_arr= np.nan_to_num(np.array(image_mask), nan=2)
    stacked_mask = np.stack((mask_arr,)*1, axis=-1)
    test_mask.append(stacked_mask)

#create tensorflow dataset 
train= tf.data.Dataset.from_tensor_slices((train_image, train_mask))
test = tf.data.Dataset.from_tensor_slices((test_img, test_mask))

#set parameters
train_length = len(train_image)
img_shape = (512,512,1)
batch_size = 16
epochs = 20

#shuffle, batch, and repeat
train_dataset = train.cache().shuffle(train_length).batch(batch_size).repeat()
train_dataset = train_dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
test_dataset = test.batch(batch_size).repeat()

#build the model
def conv_block(input_tensor, num_filters):
    encoder = layers.Conv2D(num_filters, (3, 3), padding='same')(input_tensor)
    encoder = layers.BatchNormalization()(encoder)
    encoder = layers.Activation('relu')(encoder)
    encoder = layers.Conv2D(num_filters, (3, 3), padding='same')(encoder)
    encoder = layers.BatchNormalization()(encoder)
    encoder = layers.Activation('relu')(encoder)
    return encoder

def encoder_block(input_tensor, num_filters):
    encoder = conv_block(input_tensor, num_filters)
    encoder_pool = layers.MaxPooling2D((2, 2), strides=(2, 2))(encoder)
    return encoder_pool, encoder

def decoder_block(input_tensor, concat_tensor, num_filters):
    decoder = layers.Conv2DTranspose(num_filters, (2, 2), strides=(2, 2), padding='same')(input_tensor)
    decoder = layers.concatenate([concat_tensor, decoder], axis=-1)
    decoder = layers.BatchNormalization()(decoder)
    decoder = layers.Activation('relu')(decoder)
    decoder = layers.Conv2D(num_filters, (3, 3), padding='same')(decoder)
    decoder = layers.BatchNormalization()(decoder)
    decoder = layers.Activation('relu')(decoder)
    decoder = layers.Conv2D(num_filters, (3, 3), padding='same')(decoder)
    decoder = layers.BatchNormalization()(decoder)
    decoder = layers.Activation('relu')(decoder)
    return decoder

inputs = layers.Input(shape=img_shape)
encoder0_pool, encoder0 = encoder_block(inputs, 32)
encoder1_pool, encoder1 = encoder_block(encoder0_pool, 64)
encoder2_pool, encoder2 = encoder_block(encoder1_pool, 128)
encoder3_pool, encoder3 = encoder_block(encoder2_pool, 256)
encoder4_pool, encoder4 = encoder_block(encoder3_pool, 512)
center = conv_block(encoder4_pool, 1024)
decoder4 = decoder_block(center, encoder4, 512)
decoder3 = decoder_block(decoder4, encoder3, 256)
decoder2 = decoder_block(decoder3, encoder2, 128)
decoder1 = decoder_block(decoder2, encoder1, 64)
decoder0 = decoder_block(decoder1, encoder0, 32)
outputs = layers.Conv2D(1, (1, 1), activation='sigmoid')(decoder0)

#defined the model
model = models.Model(inputs=[inputs], outputs=[outputs])

#defined loss function
def dice_coeff(y_true, y_pred):
    smooth = 1.
    y_true_f = tf.reshape(y_true, [-1])
    y_pred_f = tf.reshape(y_pred, [-1])
    intersection = tf.reduce_sum(y_true_f * y_pred_f)
    score = (2.*intersection+smooth)/(tf.reduce_sum(y_true_f)+tf.reduce_sum(y_pred_f)+smooth)
    return score

def dice_loss(y_true, y_pred):
    loss = 1 - dice_coeff(y_true, y_pred)
    return loss

def bce_dice_loss(y_true, y_pred):
    loss = losses.binary_crossentropy(y_true, y_pred) + dice_loss(y_true, y_pred)
    return loss

#compiled the model
model.compile(optimizer='adam', loss=bce_dice_loss, metrics=[dice_loss])
model.summary()

save_model_path = '/content/drive/My Drive/tmp/weights.hdf5'
cp = tf.keras.callbacks.ModelCheckpoint(filepath=save_model_path, monitor='val_dice_loss', mode='max', save_best_only=True)

#trained the model
history = model.fit(train_dataset, steps_per_epoch=int(np.ceil(train_length / float(batch_size))), epochs=epochs, validation_data=test_dataset, validation_steps=int(np.ceil(len(test_img) / float(batch_size))), callbacks=[cp])

#visualize training process
dice = history.history['dice_loss']
val_dice = history.history['val_dice_loss']

loss = history.history['loss']
val_loss = history.history['val_loss']

epochs_range = range(epochs)

plt.figure(figsize=(16, 8))
plt.subplot(1, 2, 1)
plt.plot(epochs_range, dice, label='Training Dice Loss')
plt.plot(epochs_range, val_dice, label='Validation Dice Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Dice Loss')

plt.subplot(1, 2, 2)
plt.plot(epochs_range, loss, label='Training Loss')
plt.plot(epochs_range, val_loss, label='Validation Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')

plt.show()

#visualize the output
def display(display_list):
    plt.figure(figsize=(15, 15))
    title = ['Input Image', 'True Mask', 'Predicted Mask']
    for i in range(len(display_list)):
        plt.subplot(1, len(display_list), i+1)
        plt.title(title[i])
        plt.imshow(tf.keras.preprocessing.image.array_to_img(display_list[i]))
        plt.axis('off')
    plt.show()

def show_predictions(dataset=None, num=1):
    for image, mask in dataset.take(num):
        pred_mask = model.predict(image)
        display([image[0,:,:,0], mask[0,:,:,0], create_mask(pred_mask)[:,:,0]]) #1
        display([image[0,:,:,0], mask[0,:,:,0], pred_mask[0,:,:,0]]) #2

def create_mask(pred_mask):
    pred_mask = tf.argmax(pred_mask, axis=-1)
    pred_mask = pred_mask[..., tf.newaxis]
    return pred_mask[0]

show_predictions(test_dataset, 3)

The output classifier pred_mask do show some meaningful structures, below are some examples of the output from code #1. Example 1 . Example 2 . Example 3 .

When I tried to create a segmentation mask like TensorFlow tutorial with code #2, it returned 0 all over for the segmentation mask. Example 1 .

Answer 1

I find out how to produce the mask. Below is what required to be modified.

#read images as array and make their shape
train_image = []
for m in img[:]:
    image= Image.open(m)
    img_arr= np.nan_to_num(np.array(image), nan=0)
    stacked_img= np.stack((img_arr,)*3, axis=-1)
    train_image.append(stacked_img)

test_img = []
for o in img_val[:]:
    image= Image.open(o)
    img_arr = np.nan_to_num(np.array(image), nan=0)
    stacked_img = np.stack((img_arr,)*3, axis=-1)
    test_img.append(stacked_img)

#set parameters
img_shape = (512,512,3) #if I want to produce a 3-class mask, then set the third channel as 3

#build the model
outputs = layers.Conv2D(3, (1, 1), activation='softmax')(decoder0) #if I want to produce 3-class mask, then set layers.Conv2D(3,(1,1)) and use softmax; if 2-class mask is required, then set (2,(1,1)) and use sigmoid.

Though it works now, I am confused about the image channel and the parameter layers.Conv2D. If I have a RGB image and I want to produce a 5-class mask, how should I set for the parameter img_shape as input and layers.Conv2D (?,(1,1)) as output?