This repository contains the implementation of the de-noising model described in DeepErase: Weakly Supervised Ink Artifact Removal in Document Text Images .
Our implementation also has the ability to work with Generative Adversarial Networks (GAN) framework to improve the noise removal.
import deeperase
input_pattern = "data/noised/*.jpg"
target_pattern = "data/clean/*.jpg"
image_size = (64, 256)
train_ds, test_ds = deeperase.data.build_dataset(
input_pattern, target_pattern, image_size=image_size,
do_crop=True, do_hflip=True) # Also with image augmentation 😲
generator = deeperase.build_unet(image_size + (3,),
backbone='mobilenetv2', # See deeperase/backbones.py
n_channels=3) # rgb
discriminator = deeperase.build_discriminator(image_size + (3,))
# If discriminator is not None the noise removal model will also use an
# adversarial loss
de = deeperase.DeepErase(generator, discriminator)
de.compile(optimizer=tf.optimizers.Adam(learning_rate=2e-4, beta_1=.5),
discriminator_optimizer=tf.optimizers.Adam(learning_rate=2e-4,
beta_1=.5))
de.fit(train_ds.batch(64), epochs=50,
callbacks=[deeperase.callbacks.TensorBoard(logdir='logs,
images_dataset=test_ds)])
# Take a look at Tensorboard to visualize how predictions evolve 📈