Super-Resolution

In this project we provide an implementation for single image super-resolution, a process that is able to recover high-resolution images from low-resolution inputs as demonstrated in the example images. Our system is based on the TensorFlow 2.x framework and implements several popular single image super-resolution models:

• Enhanced Deep Residual Networks for Single Image Super-Resolution (EDSR), winner of the NTIRE 2017 super-resolution challenge,
• Wide Activation for Efficient and Accurate Image Super-Resolution (WDSR), winner of the NTIRE 2018 super-resolution challenge (realistic tracks) and
• Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network (SRGAN).

The high-level training API of the project makes it easy to train these models as described in their respective papers. We also support fine-tuning EDSR and WDSR models in an SRGAN context for improved perceptual quality.

For ease of use, we have created a DIV2K data provider that automatically downloads training and validation images at various scales and with different downgrade operators (e.g. bicubic). This allows for easy evaluation of model performance on standard datasets.

This implementation is used in several projects and papers such as:

• Satellite image multi-frame super resolution using 3D wide-activation neural networks
• AutoSR4EO: An AutoML Approach to Super-Resolution for Earth Observation Images
• A new public Alsat-2B dataset for single-image super-resolution

Related articles:

• Single image super-resolution with deep neural networks

Links:

• Citations