2.1 MiB
2.1 MiB
In [1]:
import os, shutil from tensorflow import keras from tensorflow.keras import layers from tensorflow.keras import models from keras.layers import Input, Dense, Conv2D, MaxPool2D, UpSampling2D, Dropout from keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau import numpy as np import matplotlib.pyplot as plt from PIL import Image, ImageFile from PIL import ImageFont from PIL import ImageDraw, ImageOps import string
Generate input images with watermarks¶
In [2]:
# list printable characters printable = list(string.printable)[0:84]
In [3]:
def gen_watermark(img_str, train=1): # open the image as jpegs img = Image.open("/home/tensorflow/Downloads/VOCdevkit/RAWS/{}".format(img_str)) # generate random font sizes 30 - 100 np.random.seed() font_size = np.random.randint(low = 30, high = 100) # create the watermark font font = ImageFont.truetype("/home/tensorflow/Downloads/VOCdevkit/abel-regular.ttf", font_size) # generate image to hold the watermark img_temp = Image.new('L', (350,350)) # create the watermark text np.random.seed() text_str = np.random.choice(printable, np.random.randint(low=5, high=10)) text_str = "".join(text_str) # draw on the temporary image draw_temp = ImageDraw.Draw(img_temp) # generate random opacity level np.random.seed() opacity = np.random.randint(low=90, high=120) # insert text onto the temporary image draw_temp.text((0,0), text_str, font=font, fill=opacity) # generate random rotation angle np.random.seed() rot_int = np.random.randint(low=0, high=40) # rotate the text rotated_text = img_temp.rotate(rot_int, expand=1) # default text color = white col_1 = (255,255,255) col_2 = (255,255,255) # generate random location to put the text np.random.seed() rand_loc = tuple(np.random.randint(low=10,high=100,size=(2,))) # merge the temporary image img.paste(ImageOps.colorize(rotated_text, col_1, col_2), rand_loc, rotated_text) # save the watermarked image img.save("/home/tensorflow/Downloads/VOCdevkit/WATS/{}".format(img_str))
In [4]:
# Prepare the directory wat_path = "/home/tensorflow/Downloads/VOCdevkit/WATS" if not os.path.exists(wat_path): os.mkdir(wat_path) else: shutil.rmtree(wat_path)
In [5]:
# Or else it will crash ImageFile.LOAD_TRUNCATED_IMAGES = True
In [6]:
# Generate watermarked images img_list = os.listdir("/home/tensorflow/Downloads/VOCdevkit/RAWS") for _, img in enumerate(img_list): gen_watermark(img)
In [7]:
# Show some images raw_str = "/home/tensorflow/Downloads/VOCdevkit/RAWS/" wat_str = "/home/tensorflow/Downloads/VOCdevkit/WATS/" dir_raw = os.listdir("/home/tensorflow/Downloads/VOCdevkit/RAWS") dir_wat = os.listdir("/home/tensorflow/Downloads/VOCdevkit/WATS") f, axs = plt.subplots(3,2, figsize=(25, 25), sharey=True, sharex = True) axs = axs.ravel() img_raw = Image.open(raw_str + dir_raw[4]) img_wat = Image.open(wat_str + dir_wat[4]) axs[0].imshow(img_raw) axs[1].imshow(img_wat) img_raw = Image.open(raw_str + dir_raw[5]) img_wat = Image.open(wat_str + dir_wat[5]) axs[2].imshow(img_raw) axs[3].imshow(img_wat) img_raw = Image.open(raw_str + dir_raw[10]) img_wat = Image.open(wat_str + dir_wat[10]) axs[4].imshow(img_raw) axs[5].imshow(img_wat) plt.show()
Network Code¶
Transform inputs and labels¶
In [8]:
train_wat_path = "/home/tensorflow/Downloads/VOCdevkit/train/wat_imgs" train_raw_path = "/home/tensorflow/Downloads/VOCdevkit/train/raw_imgs" val_wat_path = "/home/tensorflow/Downloads/VOCdevkit/val/wat_imgs" val_raw_path = "/home/tensorflow/Downloads/VOCdevkit/val/raw_imgs"
In [9]:
# Seperate training and validation inputs/labels if not os.path.exists(train_raw_path): os.mkdir(train_raw_path) if not os.path.exists(val_raw_path): os.mkdir(val_raw_path) if not os.path.exists(train_wat_path): os.mkdir(train_wat_path) else: shutil.rmtree(train_wat_path) if not os.path.exists(val_wat_path): os.mkdir(val_wat_path) else: shutil.rmtree(val_wat_path)
In [10]:
# Copy raw images over raw_list = os.listdir("/home/tensorflow/Downloads/VOCdevkit/RAWS") for ind, img_str in enumerate(raw_list): img = Image.open("/home/tensorflow/Downloads/VOCdevkit/RAWS/{}".format(img_str)) if ind < 4900: img.save("/home/tensorflow/Downloads/VOCdevkit/train/raw_imgs/{}".format(img_str)) else: img.save("/home/tensorflow/Downloads/VOCdevkit/val/raw_imgs/{}".format(img_str))
In [11]:
# Copy watermarked images over wat_list = os.listdir("/home/tensorflow/Downloads/VOCdevkit/WATS") for ind, img_str in enumerate(wat_list): img = Image.open("/home/tensorflow/Downloads/VOCdevkit/WATS/{}".format(img_str)) if ind < 4900: img.save("/home/tensorflow/Downloads/VOCdevkit/train/wat_imgs/{}".format(img_str)) else: img.save("/home/tensorflow/Downloads/VOCdevkit/val/wat_imgs/{}".format(img_str))
In [12]:
# Standardize and normalize x_train = np.ndarray(shape=(4900, 128, 128, 3), dtype=np.float32) y_train = np.ndarray(shape=(4900, 128, 128, 3), dtype=np.float32) x_val = np.ndarray(shape=(196, 128, 128, 3), dtype=np.float32) y_val = np.ndarray(shape=(196, 128, 128, 3), dtype=np.float32) i = 0 for file in os.listdir(train_wat_path): img = Image.open("/home/tensorflow/Downloads/VOCdevkit/train/wat_imgs/{}".format(file)) img = np.array(img.resize((128,128))) / 255 img = img.astype('float32') x_train[i] = img i += 1 i = 0 for file in os.listdir(train_raw_path): img = Image.open("/home/tensorflow/Downloads/VOCdevkit/train/raw_imgs/{}".format(file)) img = np.array(img.resize((128,128))) / 255 img = img.astype('float32') y_train[i] = img i += 1 i = 0 for file in os.listdir(val_wat_path): img = Image.open("/home/tensorflow/Downloads/VOCdevkit/val/wat_imgs/{}".format(file)) img = np.array(img.resize((128,128))) / 255 img = img.astype('float32') x_val[i] = img i += 1 i = 0 for file in os.listdir(val_raw_path): img = Image.open("/home/tensorflow/Downloads/VOCdevkit/val/raw_imgs/{}".format(file)) img = np.array(img.resize((128,128))) / 255 img = img.astype('float32') y_val[i] = img i += 1
In [13]:
# See images after transforming fig, axs = plt.subplots(1, 2, figsize=(25, 25), sharey=True, sharex = True) img_1 = x_train[0] img_2 = y_train[0] axs[0].imshow(img_1) axs[1].imshow(img_2)
Out[13]:
<matplotlib.image.AxesImage at 0x7f0c99cd56d0>
Create model¶
In [14]:
class inpaintingModel: ''' Build UNET like model for image inpaining task. ''' def prepare_model(self, input_size=(128,128,3)): inputs = keras.layers.Input(input_size) conv1, pool1 = self.__ConvBlock(32, (3,3), (2,2), 'relu', 'same', inputs) pool1 = Dropout(0.25)(pool1) conv2, pool2 = self.__ConvBlock(64, (3,3), (2,2), 'relu', 'same', pool1) pool2 = Dropout(0.5)(pool2) conv3, pool3 = self.__ConvBlock(128, (3,3), (2,2), 'relu', 'same', pool2) pool3 = Dropout(0.5)(pool3) conv4, pool4 = self.__ConvBlock(256, (3,3), (2,2), 'relu', 'same', pool3) pool4 = Dropout(0.5)(pool4) conv5, up6 = self.__UpConvBlock(512, 256, (3,3), (2,2), (2,2), 'relu', 'same', pool4, conv4) up6 = Dropout(0.5)(up6) conv6, up7 = self.__UpConvBlock(256, 128, (3,3), (2,2), (2,2), 'relu', 'same', up6, conv3) up7 = Dropout(0.5)(up7) conv7, up8 = self.__UpConvBlock(128, 64, (3,3), (2,2), (2,2), 'relu', 'same', up7, conv2) up8= Dropout(0.5)(up8) conv8, up9 = self.__UpConvBlock(64, 32, (3,3), (2,2), (2,2), 'relu', 'same', up8, conv1) up9= Dropout(0.25)(up9) conv9 = self.__ConvBlock(32, (3,3), (2,2), 'relu', 'same', up9, False) outputs = keras.layers.Conv2D(3, (3, 3), activation='sigmoid', padding='same')(conv9) return keras.models.Model(inputs=[inputs], outputs=[outputs]) def __ConvBlock(self, filters, kernel_size, pool_size, activation, padding, connecting_layer, pool_layer=True): conv = keras.layers.Conv2D(filters=filters, kernel_size=kernel_size, activation=activation, padding=padding)(connecting_layer) conv = keras.layers.Conv2D(filters=filters, kernel_size=kernel_size, activation=activation, padding=padding)(conv) if pool_layer: pool = keras.layers.MaxPooling2D(pool_size)(conv) return conv, pool else: return conv def __UpConvBlock(self, filters, up_filters, kernel_size, up_kernel, up_stride, activation, padding, connecting_layer, shared_layer): conv = keras.layers.Conv2D(filters=filters, kernel_size=kernel_size, activation=activation, padding=padding)(connecting_layer) conv = keras.layers.Conv2D(filters=filters, kernel_size=kernel_size, activation=activation, padding=padding)(conv) up = keras.layers.Conv2DTranspose(filters=up_filters, kernel_size=up_kernel, strides=up_stride, padding=padding)(conv) up = keras.layers.concatenate([up, shared_layer], axis=3) return conv, up
In [15]:
model = inpaintingModel().prepare_model() model.compile(loss="binary_crossentropy", optimizer="adam", metrics=["accuracy"]) model.summary()
Model: "model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 128, 128, 3) 0
__________________________________________________________________________________________________
conv2d (Conv2D) (None, 128, 128, 32) 896 input_1[0][0]
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 128, 128, 32) 9248 conv2d[0][0]
__________________________________________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 64, 64, 32) 0 conv2d_1[0][0]
__________________________________________________________________________________________________
dropout (Dropout) (None, 64, 64, 32) 0 max_pooling2d[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 64, 64, 64) 18496 dropout[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 64, 64, 64) 36928 conv2d_2[0][0]
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None, 32, 32, 64) 0 conv2d_3[0][0]
__________________________________________________________________________________________________
dropout_1 (Dropout) (None, 32, 32, 64) 0 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 32, 32, 128) 73856 dropout_1[0][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D) (None, 32, 32, 128) 147584 conv2d_4[0][0]
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D) (None, 16, 16, 128) 0 conv2d_5[0][0]
__________________________________________________________________________________________________
dropout_2 (Dropout) (None, 16, 16, 128) 0 max_pooling2d_2[0][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D) (None, 16, 16, 256) 295168 dropout_2[0][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D) (None, 16, 16, 256) 590080 conv2d_6[0][0]
__________________________________________________________________________________________________
max_pooling2d_3 (MaxPooling2D) (None, 8, 8, 256) 0 conv2d_7[0][0]
__________________________________________________________________________________________________
dropout_3 (Dropout) (None, 8, 8, 256) 0 max_pooling2d_3[0][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D) (None, 8, 8, 512) 1180160 dropout_3[0][0]
__________________________________________________________________________________________________
conv2d_9 (Conv2D) (None, 8, 8, 512) 2359808 conv2d_8[0][0]
__________________________________________________________________________________________________
conv2d_transpose (Conv2DTranspo (None, 16, 16, 256) 524544 conv2d_9[0][0]
__________________________________________________________________________________________________
concatenate (Concatenate) (None, 16, 16, 512) 0 conv2d_transpose[0][0]
conv2d_7[0][0]
__________________________________________________________________________________________________
dropout_4 (Dropout) (None, 16, 16, 512) 0 concatenate[0][0]
__________________________________________________________________________________________________
conv2d_10 (Conv2D) (None, 16, 16, 256) 1179904 dropout_4[0][0]
__________________________________________________________________________________________________
conv2d_11 (Conv2D) (None, 16, 16, 256) 590080 conv2d_10[0][0]
__________________________________________________________________________________________________
conv2d_transpose_1 (Conv2DTrans (None, 32, 32, 128) 131200 conv2d_11[0][0]
__________________________________________________________________________________________________
concatenate_1 (Concatenate) (None, 32, 32, 256) 0 conv2d_transpose_1[0][0]
conv2d_5[0][0]
__________________________________________________________________________________________________
dropout_5 (Dropout) (None, 32, 32, 256) 0 concatenate_1[0][0]
__________________________________________________________________________________________________
conv2d_12 (Conv2D) (None, 32, 32, 128) 295040 dropout_5[0][0]
__________________________________________________________________________________________________
conv2d_13 (Conv2D) (None, 32, 32, 128) 147584 conv2d_12[0][0]
__________________________________________________________________________________________________
conv2d_transpose_2 (Conv2DTrans (None, 64, 64, 64) 32832 conv2d_13[0][0]
__________________________________________________________________________________________________
concatenate_2 (Concatenate) (None, 64, 64, 128) 0 conv2d_transpose_2[0][0]
conv2d_3[0][0]
__________________________________________________________________________________________________
dropout_6 (Dropout) (None, 64, 64, 128) 0 concatenate_2[0][0]
__________________________________________________________________________________________________
conv2d_14 (Conv2D) (None, 64, 64, 64) 73792 dropout_6[0][0]
__________________________________________________________________________________________________
conv2d_15 (Conv2D) (None, 64, 64, 64) 36928 conv2d_14[0][0]
__________________________________________________________________________________________________
conv2d_transpose_3 (Conv2DTrans (None, 128, 128, 32) 8224 conv2d_15[0][0]
__________________________________________________________________________________________________
concatenate_3 (Concatenate) (None, 128, 128, 64) 0 conv2d_transpose_3[0][0]
conv2d_1[0][0]
__________________________________________________________________________________________________
dropout_7 (Dropout) (None, 128, 128, 64) 0 concatenate_3[0][0]
__________________________________________________________________________________________________
conv2d_16 (Conv2D) (None, 128, 128, 32) 18464 dropout_7[0][0]
__________________________________________________________________________________________________
conv2d_17 (Conv2D) (None, 128, 128, 32) 9248 conv2d_16[0][0]
__________________________________________________________________________________________________
conv2d_18 (Conv2D) (None, 128, 128, 3) 867 conv2d_17[0][0]
==================================================================================================
Total params: 7,760,931
Trainable params: 7,760,931
Non-trainable params: 0
__________________________________________________________________________________________________
Training¶
In [16]:
early_stopping = EarlyStopping(patience=10, verbose=1) model_checkpoint = ModelCheckpoint("./keras.model", save_best_only=True, verbose=1) reduce_lr = ReduceLROnPlateau(factor=0.1, patience=5, min_lr=0.00001, verbose=1) history = model.fit(x_train, y_train, validation_data=(x_val, y_val), epochs=30, batch_size=32, callbacks=[early_stopping, model_checkpoint, reduce_lr])
Epoch 1/30 154/154 [==============================] - 132s 764ms/step - loss: 0.5882 - accuracy: 0.5445 - val_loss: 0.5275 - val_accuracy: 0.8164 Epoch 00001: val_loss improved from inf to 0.52747, saving model to ./keras.model INFO:tensorflow:Assets written to: ./keras.model/assets Epoch 2/30 154/154 [==============================] - 113s 735ms/step - loss: 0.5266 - accuracy: 0.7876 - val_loss: 0.5246 - val_accuracy: 0.8419 Epoch 00002: val_loss improved from 0.52747 to 0.52456, saving model to ./keras.model INFO:tensorflow:Assets written to: ./keras.model/assets Epoch 3/30 154/154 [==============================] - 114s 739ms/step - loss: 0.5215 - accuracy: 0.8043 - val_loss: 0.5226 - val_accuracy: 0.8693 Epoch 00003: val_loss improved from 0.52456 to 0.52256, saving model to ./keras.model INFO:tensorflow:Assets written to: ./keras.model/assets Epoch 4/30 154/154 [==============================] - 114s 738ms/step - loss: 0.5216 - accuracy: 0.8182 - val_loss: 0.5210 - val_accuracy: 0.8521 Epoch 00004: val_loss improved from 0.52256 to 0.52101, saving model to ./keras.model INFO:tensorflow:Assets written to: ./keras.model/assets Epoch 5/30 154/154 [==============================] - 114s 740ms/step - loss: 0.5205 - accuracy: 0.8236 - val_loss: 0.5200 - val_accuracy: 0.8616 Epoch 00005: val_loss improved from 0.52101 to 0.52001, saving model to ./keras.model INFO:tensorflow:Assets written to: ./keras.model/assets Epoch 6/30 154/154 [==============================] - 114s 742ms/step - loss: 0.5196 - accuracy: 0.8239 - val_loss: 0.5194 - val_accuracy: 0.8534 Epoch 00006: val_loss improved from 0.52001 to 0.51936, saving model to ./keras.model INFO:tensorflow:Assets written to: ./keras.model/assets Epoch 7/30 154/154 [==============================] - 114s 744ms/step - loss: 0.5193 - accuracy: 0.8306 - val_loss: 0.5199 - val_accuracy: 0.8710 Epoch 00007: val_loss did not improve from 0.51936 Epoch 8/30 154/154 [==============================] - 115s 747ms/step - loss: 0.5190 - accuracy: 0.8328 - val_loss: 0.5202 - val_accuracy: 0.8559 Epoch 00008: val_loss did not improve from 0.51936 Epoch 9/30 154/154 [==============================] - 115s 746ms/step - loss: 0.5192 - accuracy: 0.8367 - val_loss: 0.5195 - val_accuracy: 0.8664 Epoch 00009: val_loss did not improve from 0.51936 Epoch 10/30 154/154 [==============================] - 115s 748ms/step - loss: 0.5188 - accuracy: 0.8373 - val_loss: 0.5198 - val_accuracy: 0.8769 Epoch 00010: val_loss did not improve from 0.51936 Epoch 11/30 154/154 [==============================] - 115s 748ms/step - loss: 0.5163 - accuracy: 0.8358 - val_loss: 0.5190 - val_accuracy: 0.8400 Epoch 00011: val_loss improved from 0.51936 to 0.51905, saving model to ./keras.model INFO:tensorflow:Assets written to: ./keras.model/assets Epoch 12/30 154/154 [==============================] - 115s 744ms/step - loss: 0.5186 - accuracy: 0.8335 - val_loss: 0.5193 - val_accuracy: 0.8613 Epoch 00012: val_loss did not improve from 0.51905 Epoch 13/30 154/154 [==============================] - 115s 745ms/step - loss: 0.5177 - accuracy: 0.8381 - val_loss: 0.5185 - val_accuracy: 0.8720 Epoch 00013: val_loss improved from 0.51905 to 0.51849, saving model to ./keras.model INFO:tensorflow:Assets written to: ./keras.model/assets Epoch 14/30 154/154 [==============================] - 115s 747ms/step - loss: 0.5185 - accuracy: 0.8377 - val_loss: 0.5188 - val_accuracy: 0.8541 Epoch 00014: val_loss did not improve from 0.51849 Epoch 15/30 154/154 [==============================] - 115s 747ms/step - loss: 0.5168 - accuracy: 0.8358 - val_loss: 0.5215 - val_accuracy: 0.8176 Epoch 00015: val_loss did not improve from 0.51849 Epoch 16/30 154/154 [==============================] - 115s 748ms/step - loss: 0.5181 - accuracy: 0.8273 - val_loss: 0.5177 - val_accuracy: 0.8593 Epoch 00016: val_loss improved from 0.51849 to 0.51772, saving model to ./keras.model INFO:tensorflow:Assets written to: ./keras.model/assets Epoch 17/30 154/154 [==============================] - 116s 750ms/step - loss: 0.5159 - accuracy: 0.8350 - val_loss: 0.5185 - val_accuracy: 0.8332 Epoch 00017: val_loss did not improve from 0.51772 Epoch 18/30 154/154 [==============================] - 115s 747ms/step - loss: 0.5184 - accuracy: 0.8389 - val_loss: 0.5183 - val_accuracy: 0.8424 Epoch 00018: val_loss did not improve from 0.51772 Epoch 19/30 154/154 [==============================] - 115s 746ms/step - loss: 0.5178 - accuracy: 0.8434 - val_loss: 0.5183 - val_accuracy: 0.8688 Epoch 00019: val_loss did not improve from 0.51772 Epoch 20/30 154/154 [==============================] - 115s 747ms/step - loss: 0.5170 - accuracy: 0.8327 - val_loss: 0.5182 - val_accuracy: 0.8720 Epoch 00020: val_loss did not improve from 0.51772 Epoch 21/30 154/154 [==============================] - 115s 747ms/step - loss: 0.5168 - accuracy: 0.8462 - val_loss: 0.5183 - val_accuracy: 0.8682 Epoch 00021: val_loss did not improve from 0.51772 Epoch 00021: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513. Epoch 22/30 154/154 [==============================] - 115s 749ms/step - loss: 0.5172 - accuracy: 0.8465 - val_loss: 0.5179 - val_accuracy: 0.8762 Epoch 00022: val_loss did not improve from 0.51772 Epoch 23/30 154/154 [==============================] - 116s 750ms/step - loss: 0.5179 - accuracy: 0.8504 - val_loss: 0.5180 - val_accuracy: 0.8705 Epoch 00023: val_loss did not improve from 0.51772 Epoch 24/30 154/154 [==============================] - 115s 747ms/step - loss: 0.5171 - accuracy: 0.8483 - val_loss: 0.5179 - val_accuracy: 0.8772 Epoch 00024: val_loss did not improve from 0.51772 Epoch 25/30 154/154 [==============================] - 115s 748ms/step - loss: 0.5169 - accuracy: 0.8478 - val_loss: 0.5181 - val_accuracy: 0.8749 Epoch 00025: val_loss did not improve from 0.51772 Epoch 26/30 154/154 [==============================] - 115s 748ms/step - loss: 0.5177 - accuracy: 0.8507 - val_loss: 0.5179 - val_accuracy: 0.8740 Epoch 00026: val_loss did not improve from 0.51772 Epoch 00026: ReduceLROnPlateau reducing learning rate to 1.0000000474974514e-05. Epoch 00026: early stopping
In [17]:
# Visualize the training process fig, (ax_loss, ax_acc) = plt.subplots(1, 2, figsize=(15,5)) ax_loss.plot(history.epoch, history.history["loss"], label="Train loss") ax_loss.plot(history.epoch, history.history["val_loss"], label="Validation loss") ax_acc.plot(history.epoch, history.history["accuracy"], label="Train accuracy") ax_acc.plot(history.epoch, history.history["val_accuracy"], label="Validation accuracy")
Out[17]:
[<matplotlib.lines.Line2D at 0x7f0d0cb4ef70>]
Test¶
In [ ]:
model = load_model("./keras.model")
In [18]:
# Generate watermark (again) def watermark(img_str): img = Image.open("/home/tensorflow/Downloads/VOCdevkit/TEST/RAW/{}".format(img_str)) img = img.resize((512,round(img.size[1]*512/img.size[0]))) np.random.seed() font_size = np.random.randint(low = 30, high = 100) font = ImageFont.truetype("/home/tensorflow/Downloads/VOCdevkit/abel-regular.ttf", font_size) img_temp = Image.new('L', (350,350)) np.random.seed() text_str = np.random.choice(printable, np.random.randint(low=5, high=10)) text_str = "".join(text_str) draw_temp = ImageDraw.Draw(img_temp) np.random.seed() opacity = np.random.randint(low=90, high=120) # insert text onto the temporary image draw_temp.text((0,0), text_str, font=font, fill=opacity) np.random.seed() rot_int = np.random.randint(low=0, high=30) rotated_text = img_temp.rotate(rot_int, expand=1) col_1 = (255,255,255) col_2 = (255,255,255) np.random.seed() rand_loc = tuple(np.random.randint(low=10,high=100,size=(2,))) img.paste(ImageOps.colorize(rotated_text, col_1, col_2), rand_loc, rotated_text) img.save("/home/tensorflow/Downloads/VOCdevkit/TEST/WAT/{}".format(img_str)) return img
In [27]:
fig, axs = plt.subplots(1, 3, figsize=(25, 25), sharey=True, sharex = True) # The watermarked image wat_img = watermark("1.jpg") wat_img = np.array(wat_img.resize((128,128))) / 255 wat_img = wat_img.astype('float32') # The predicted image rec_img = model.predict(wat_img.reshape((1,)+wat_img.shape)) rec_img = rec_img.reshape(rec_img.shape[1:]) # The original image (reshaped) tr_img = Image.open("/home/tensorflow/Downloads/VOCdevkit/TEST/RAW/1.jpg") tr_img = np.array(tr_img.resize((128,128))) / 255 tr_img = tr_img.astype('float32') axs[0].imshow(rec_img) axs[1].imshow(wat_img) axs[2].imshow(tr_img)
Out[27]:
<matplotlib.image.AxesImage at 0x7f0d0c947ee0>