Image Super-Resolution GANs

Name: Image Super-Resolution GANs
Rating: 4.2 (32 reviews)

Enhance/upsample images with Generative Adversarial Networks using Python and Tensorflow 2.0

Created byBrad Klingensmith

Last updated 1/2022

English

What you'll learn

Create a generator architecture that upsamples an image by 4 times in each dimension
Create a discriminator architecture that scores both realism and fidelity to the original image
Modify custom written Keras layers to accept input images of any size without rebuilding the model
Train the models on a Cloud TPU through Google CoLab
Use the trained generator in a practical application to upsample your own images

Course content

4 sections • 24 lectures • 2h 31m total length

Introduction0:50
Prerequisites1:01
Getting the Most Out of This Course1:02
What is Super-resolution?2:27

Generator - High Level5:04
Explore the generator for image super-resolution by replacing latent noise with a low-resolution patch, using a five-block image comprehension module with residual connections, and fourfold upsampling from 64x64 to 256x256.
Generator - Handling a Variety of Image Sizes4:39
Learn to build a fully convolutional generator that is input-size agnostic, handles images of varying sizes by sharing weights across patches and training on multiple resolutions.
Generator - Details5:58
Discriminator - High Level6:29
Discriminator - Details1:55
Generator - Code5:47
Explore the generator code for image super-resolution, detailing base and super resolution handling, variably sized inputs, upsampling, and residual blocks with a 1x1 conversion to adjust the channel count.
Handling Various Image Sizes - Code3:17
Learn how image super-resolution GANs handle varying image sizes with dynamic shapes and transpose convolution. See how shape computation is deferred until runtime for TPU building and compilation.
Discriminator - Code3:09
describes a two-input discriminator for image super-resolution, using the super image and bass image, with a recursive body that downsamples, applies convolutions, and residual connections to produce a realism score.

Loss Function10:54
Keep the loss function as is, using the R2 loss with regularization to produce sharp, realistic images from the low-resolution input, avoiding content loss and relying on the discriminator.
Training Overview8:58
Train the generator and discriminator in alternating steps, using high-resolution, low-resolution, and super-resolution images. Employ the R2 loss, separate optimizers, and patch-based crops across multiple resolutions.
Training Loop Code7:14
Training Loop Code - Resizing and Cropping6:42
Explore the training loop’s resizing and cropping, featuring gamma correction and decoding, anti-aliasing, and random crop, with downsampling via average pooling to produce realistic base images.
Training Setup Code14:13
Configure the image super-resolution GAN training pipeline by moving flip augmentation to the TPU, compiling multiple datasets, and generating visualizer ready super images via a random crop-and-resize pipeline.
Visualizer Code7:48
Main Training Script and Colab Notebook8:41
Visualizations13:55

Requirements

My "High Resolution Generative Adversarial Networks (GANs)" course
Python experience
Convolutional neural network experience
Basic familiarity with TensorFlow 2.0 and Keras

Description

We've all seen the gimmick in crime TV shows where the investigators manage to take a tiny patch of an image and magnify it with unrealistic clarity. Well today, Generative Adversarial Networks are making the impossible possible.

Dive into this course where I’ll show you how easily we can take the fundamentals from my High Resolution Generative Adversarial Networks course and build on this to accomplish this impressive feat known as Super-resolution. Not only will you be able to train a Generator to magnify an image to 4 times it’s original size (that’s 16 times the number of pixel!), but it will take relatively little effort on our end.

Just as in the first course, we’ll use Python and TensorFlow 2.0 along with Keras to build and train our convolutional neural networks. And since training our networks will require a ton of computational power, we’ll once again use Google CoLab to connect to a free Cloud TPU. This will allow us to complete the training in just a few days without spending anything on hardware!

If this sounds enticing, take a few minutes to watch the free preview of the “Results!” lesson. I have no doubt that you will come away impressed.

Who this course is for:

Python + TensorFlow 2.0 developers who want to enlarge images with photorealistic detail and clarity

Image Super-Resolution GANs

What you'll learn

Explore related topics

Course content

Introduction4 lectures • 5min

Model Architecture8 lectures • 36min

Training8 lectures • 1hr 18min

Super-resolution in Action!4 lectures • 31min

Requirements

Description

Who this course is for: