Single Image Super-Resolution with SRGAN

Single Image Super-Resolution build on TensorFlow 2.11.0.

The code is based on the following paper,

• Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

---

Results

After some epochs this was the result (dataset from Set 5):

How to use

In a few simple lines you can train the model,

from srgan import SRGAN

# Instantiate SRGAN model
srgan = SRGAN()

# Define optimizers and loss function
g_optimizer = tf.keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5)
d_optimizer = tf.keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5)
loss_fn = tf.keras.losses.BinaryCrossentropy()

# Compile the model
srgan.compile(g_optimizer=g_optimizer, d_optimizer=d_optimizer, loss_fn=loss_fn)

# Train the model
srgan.fit(dataset, epochs=150)