深度學(xué)習(xí)項(xiàng)目示例 | 手把手教你使用自編碼器進(jìn)行模糊圖像修復(fù)

來(lái)源：DeepHub IMBA
本文約2600字，建議閱讀9分鐘
本文教你如何應(yīng)用深度學(xué)習(xí)處理模糊圖像。

數(shù)據(jù)集

編寫(xiě)代碼

import numpy as npimport pandas as pdimport matplotlib.pyplot as plt
%matplotlib inlineimport randomimport cv2import osimport tensorflow as tffrom tqdm import tqdm

good_frames = '/content/drive/MyDrive/mini_clean'bad_frames = '/content/drive/MyDrive/mini_blur'

clean_frames = []for file in tqdm(sorted(os.listdir(good_frames))):if any(extension in file for extension in ['.jpg', 'jpeg', '.png']):  image = tf.keras.preprocessing.image.load_img(good_frames + '/' + file, target_size=(128,128))  image = tf.keras.preprocessing.image.img_to_array(image).astype('float32') / 255  clean_frames.append(image)
clean_frames = np.array(clean_frames)
blurry_frames = []for file in tqdm(sorted(os.listdir(bad_frames))):if any(extension in file for extension in ['.jpg', 'jpeg', '.png']):  image = tf.keras.preprocessing.image.load_img(bad_frames + '/' + file, target_size=(128,128))  image = tf.keras.preprocessing.image.img_to_array(image).astype('float32') / 255  blurry_frames.append(image)
blurry_frames = np.array(blurry_frames)

from keras.layers import Dense, Inputfrom keras.layers import Conv2D, Flattenfrom keras.layers import Reshape, Conv2DTransposefrom keras.models import Modelfrom keras.callbacks import ReduceLROnPlateau, ModelCheckpointfrom keras.utils.vis_utils import plot_modelfrom keras import backend as K
random.seed = 21np.random.seed = seed

x = clean_frames;y = blurry_frames;
from sklearn.model_selection import train_test_splitx_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.2, random_state=42)

檢查訓(xùn)練和測(cè)試數(shù)據(jù)集的形狀。

print(x_train[0].shape)print(y_train[0].shape)

r = random.randint(0, len(clean_frames)-1)print(r)fig = plt.figure()fig.subplots_adjust(hspace=0.1, wspace=0.2)ax = fig.add_subplot(1, 2, 1)ax.imshow(clean_frames[r])ax = fig.add_subplot(1, 2, 2)ax.imshow(blurry_frames[r])

# Network Parametersinput_shape = (128, 128, 3)batch_size = 32kernel_size = 3latent_dim = 256
# Encoder/Decoder number of CNN layers and filters per layerlayer_filters = [64, 128, 256]

inputs = Input(shape = input_shape, name = 'encoder_input')x = inputs

for filters in layer_filters:  x = Conv2D(filters=filters,              kernel_size=kernel_size,              strides=2,              activation='relu',              padding='same')(x)shape = K.int_shape(x)x = Flatten()(x)latent = Dense(latent_dim, name='latent_vector')(x)

encoder = Model(inputs, latent, name='encoder')encoder.summary()

latent_inputs = Input(shape=(latent_dim,), name='decoder_input')x = Dense(shape[1]*shape[2]*shape[3])(latent_inputs)x = Reshape((shape[1], shape[2], shape[3]))(x)for filters in layer_filters[::-1]:  x = Conv2DTranspose(filters=filters,                      kernel_size=kernel_size,                      strides=2,                      activation='relu',                      padding='same')(x)
outputs = Conv2DTranspose(filters=3,                        kernel_size=kernel_size,                        activation='sigmoid',                        padding='same',                        name='decoder_output')(x)

decoder = Model(latent_inputs, outputs, name='decoder')decoder.summary()

autoencoder = Model(inputs, decoder(encoder(inputs)), name='autoencoder')autoencoder.summary()

autoencoder.compile(loss='mse', optimizer='adam',metrics=["acc"])

我選擇損失函數(shù)為均方誤差，優(yōu)化器為adam，評(píng)估指標(biāo)為準(zhǔn)確率。然后還需要定義學(xué)習(xí)率調(diào)整的計(jì)劃，這樣可以在指標(biāo)沒(méi)有改進(jìn)的情況下降低學(xué)習(xí)率：

lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1),                              cooldown=0,                              patience=5,                              verbose=1,                              min_lr=0.5e-6)

callbacks = [lr_reducer]

history = autoencoder.fit(blurry_frames,                    clean_frames,                    validation_data=(blurry_frames, clean_frames),                    epochs=100,                    batch_size=batch_size,                    callbacks=callbacks)

最后結(jié)果

print("\n       Input                       Ground Truth                 Predicted Value")
for i in range(3):
  r = random.randint(0, len(clean_frames)-1)
  x, y = blurry_frames[r],clean_frames[r]  x_inp=x.reshape(1,128,128,3)  result = autoencoder.predict(x_inp)  result = result.reshape(128,128,3)
  fig = plt.figure(figsize=(12,10))  fig.subplots_adjust(hspace=0.1, wspace=0.2)
  ax = fig.add_subplot(1, 3, 1)  ax.imshow(x)
  ax = fig.add_subplot(1, 3, 2)  ax.imshow(y)
  ax = fig.add_subplot(1, 3, 3)  plt.imshow(result)

plt.figure(figsize=(12,8))plt.plot(history.history['loss'])plt.plot(history.history['val_loss'])plt.legend(['Train', 'Test'])plt.xlabel('Epoch')plt.ylabel('Loss')plt.xticks(np.arange(0, 101, 25))plt.show()

plt.figure(figsize=(12,8))plt.plot(history.history['acc'])plt.plot(history.history['val_acc'])plt.legend(['Train', 'Test'])plt.xlabel('Epoch')plt.ylabel('Accuracy')plt.xticks(np.arange(0, 101, 25))plt.show()

總結(jié)