Single Image Super Resolution

Image super-resolution (SR) is the process of recovering high-resolution (HR) images from low-resolution (LR) images. It is an important class of image processing techniques in computer vision and image processing and enjoys a wide range of real-world applications, such as medical imaging, satellite imaging, surveillance and security, astronomical imaging, amongst others.

With the advancement in deep learning techniques in recent years, deep learning-based SR models have been actively explored and often achieve state-of-the-art performance on various benchmarks of SR. A variety of deep learning methods have been applied to solve SR tasks, ranging from the early Convolutional Neural Networks (CNN) based method to recent promising Generative Adversarial Nets based SR approaches.

this project was based on these two papers

Enhanced Deep Residual Networks for Single Image Super-Resolution:

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

SRResNet: adopts the design of ResNet to solve issues with training very deep models. it managed to achieve state-of-the-art performance when it came out. it contains 16 residual blocks and uses mean squared error as a loss function Here’s an overview of the SRResNet architecture:

EDSR: One super-resolution model that follows this high-level architecture is described in the paper Enhanced Deep Residual Networks for Single Image Super-Resolution (EDSR). It is a winner of the NTIRE 2017 super-resolution challenge. They further improved the performance by employing a better ResNet structure: Batch Normalization layers are removed, and instead of mean squared error, mean absolute error is used as a loss function. Here’s an overview of the EDSR architecture

SRGAN further improves the results of EDSR by fine-tuning its weights so that it can generate high-frequency details in the generated image. This is done by training the model in a GAN using the Perceptual loss function.